Monocular vision-based unmanned vehicle high-precision relative pose estimation docking method, device, equipment and medium

Abstract

This application provides a method, apparatus, device, and medium for high-precision relative pose estimation and docking of unmanned vehicles based on monocular vision. The method includes: acquiring a monocular vision image containing the docking target region; performing multi-level edge and multi-level texture feature extraction on the monocular vision image to obtain multi-scale edge pixel clusters and local texture gradient features; reconstructing the edge pixel clusters in three-dimensional space to obtain an initial point cloud structure; performing pose analysis on the point cloud structure using texture features to obtain an initial relative pose reference sequence; extracting motion trajectory curvature features from historical continuous frame pose state data and performing spatiotemporal kinematic compensation processing to generate the target relative pose result and converting it into a chassis traction control signal to execute docking. This application can significantly improve the pose estimation accuracy and robustness during unmanned vehicle docking.

Description

Technical Field

[0001] This application relates to the field of autonomous driving and intelligent robot control technology, and more specifically, to a method, device, electronic device and computer-readable storage medium for high-precision relative pose estimation and docking of unmanned vehicles based on monocular vision. Background Technology

[0002] With the acceleration of urbanization and the development of intelligent manufacturing, intelligent inspection robots with autonomous navigation capabilities are widely used in industrial scenarios (such as power stations and chemical plants) and public safety scenarios (such as parks and industrial parks) [cite: 1]. To achieve 24-hour unmanned inspection, the inspection robot has a built-in power monitoring system. When the power level is lower than a set threshold, the robot needs to autonomously return to the charging base station for precise docking and automatic charging [cite: 1]. In this process, high-precision relative pose estimation between the vehicle and the target base station is the core prerequisite for successful docking.

[0003] In existing autonomous docking solutions for unmanned vehicles or inspection robots, a static docking control architecture based on a conventional two-dimensional vision calibration board is typically adopted. This solution first captures a two-dimensional reference calibration pattern on the base station using an onboard camera; then, it uses image processing algorithms to extract the geometric center pixel coordinates of the calibration pattern; finally, based on the static image deviation between these two-dimensional pixel coordinates and the preset ideal center coordinates, it directly generates steering and straight-line control commands and sends them to the chassis motor to guide the vehicle toward the charging base station and complete physical contact.

[0004] However, this docking scheme based on two-dimensional static features has significant technical drawbacks. Due to the complex changes in lighting conditions in the docking environment and the drastic change in perspective when the vehicle approaches the base station, the extraction of a single two-dimensional image is easily affected by local noise and blurred edges, making it difficult to construct an accurate three-dimensional spatial reference. At the same time, relying solely on the static features of the current single frame image for control completely ignores the dynamic trajectory curvature changes of the vehicle itself during the approach to the target. When the vehicle is approaching in a non-linear manner or in a large curvature turning state, it is difficult to make smooth spatiotemporal matching adjustments to the chassis actuators, resulting in a large error in the relative pose estimation of the docking end, which easily leads to docking failure or equipment collisions caused by excessive chassis traction. Summary of the Invention

[0005] This application provides a method, apparatus, device, and medium for high-precision relative pose estimation and docking of unmanned vehicles based on monocular vision, so as to at least alleviate the above-mentioned technical problems.

[0006] A high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision includes the following steps: Acquire monocular visual images containing the docking target area collected by a monocular visual perception device installed at the front of the unmanned vehicle; Perform multi-level edge feature extraction processing on the monocular vision image to obtain a multi-scale edge pixel cluster of the docking target region; Multi-level texture feature extraction processing is performed on the monocular vision image to obtain local texture gradient features; The multi-scale edge pixel cluster is subjected to three-dimensional spatial reconstruction processing to obtain an initial three-dimensional point cloud structure; The initial three-dimensional point cloud structure is subjected to pose analysis processing using the local texture gradient features to obtain an initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system. Extract motion trajectory curvature features from continuous frame pose state data of autonomous vehicles within a historical sampling period; The initial relative pose reference sequence is processed by spatiotemporal kinematic compensation using the curvature features of the motion trajectory to generate the target relative pose result and convert it into a chassis traction control signal, so that the chassis actuator can perform a matching docking control action according to the chassis traction control signal.

[0007] Optionally, the steps of performing multi-level edge feature extraction processing on the monocular vision image to obtain multi-scale edge pixel clusters of the docking target region, and performing multi-level texture feature extraction processing on the monocular vision image to obtain local texture gradient features include: The monocular visual image is mapped from the reference three primary color color space to the hue, saturation and brightness color space to obtain the converted image data; A pre-configured spatial smoothing mask is invoked to perform noise filtering on the transformed image data, resulting in smoothed image data; Extract the pixel brightness abrupt change regions from the smoothed image data; The pixel brightness abrupt change regions are aggregated to perform multi-level edge feature extraction processing, thereby obtaining a multi-scale edge pixel cluster of the docking target region; Extract the grayscale change rate between adjacent pixels in the smoothed image data to generate a basic texture gradient field; The smoothed image data is divided into multiple local spatial blocks of different scales; Within the multiple local spatial blocks of different scales, the direction and amplitude of the basic texture gradient field are statistically aggregated to perform multi-level texture feature extraction processing, thereby obtaining the local texture gradient features.

[0008] Optionally, the step of performing three-dimensional spatial reconstruction processing on the multi-scale edge pixel cluster to obtain an initial three-dimensional point cloud structure; and performing pose analysis processing on the initial three-dimensional point cloud structure using the local texture gradient features to obtain an initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system includes: Obtain the pre-calibrated set of internal calibration parameters of the monocular vision sensing device; The internal calibration parameter set is used to perform inverse projection calculation on the two-dimensional pixel coordinates in the multi-scale edge pixel cluster to generate a first three-dimensional spatial ray set; Calculate the spatial intersection points of the first three-dimensional spatial ray set within the preset depth constraint range; Spatial topology aggregation is performed on the spatial intersections to perform the three-dimensional spatial reconstruction process, resulting in the initial three-dimensional point cloud structure. The estimated normal vector of the docking target surface is determined based on the local texture gradient features. The initial 3D point cloud structure is projected onto the projection plane corresponding to the normal vector estimate to obtain the relative translation vector and the relative rotation feature set. The pose analysis process is performed based on the relative translation vector and the relative rotation feature set to obtain the initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system.

[0009] Optionally, the step of extracting motion trajectory curvature features from the pose state data of the unmanned vehicle in continuous frames within the historical sampling period includes: The inertial measurement unit inside the unmanned vehicle records the acceleration time series and angular velocity time series during the historical sampling period. The continuous frame pose state data includes the acceleration time series and angular velocity time series. Perform a first time accumulation operation on the acceleration time series to obtain the velocity time series; Perform a second time accumulation operation on the velocity time series to obtain the displacement time series; Based on the displacement time series and the angular velocity time series, the continuous frame pose state data is generated; Determine the pose points of three adjacent frames in the continuous frame pose state data; Calculate the spatial dimensions of the circumcircle of the spatial triangle formed by the pose points of the three adjacent frames; The curvature characteristics of the motion trajectory are calculated based on the circumscribed circle space dimensions.

[0010] Optionally, the step of performing spatiotemporal kinematic compensation processing on the initial relative pose reference sequence using the motion trajectory curvature features to generate the target relative pose result includes: Based on the curvature characteristics of the motion trajectory, determine whether the current motion state of the unmanned vehicle is in a high curvature turning state; When the judgment result indicates that the motion trajectory is in the large curvature turning state, a first weight adjustment coefficient is generated based on the curvature characteristics of the motion trajectory. The lateral offset in the initial relative pose reference sequence is amplified using the first weight adjustment coefficient to obtain transition pose data; Extract the chassis speed feedback quantity output in real time from the chassis speed measuring instrument of the unmanned vehicle to construct a state observation data structure; The transition pose data is introduced as an observation input into the state observation data structure for iterative convergence processing to perform spatiotemporal kinematic compensation processing, thereby generating the target relative pose result.

[0011] Optionally, the step of converting it into a chassis traction control signal includes: The target relative pose result is analyzed to separate the relative distance deviation and relative heading angle deviation between the current position of the unmanned vehicle and the target docking position; A basic longitudinal traction force value is generated based on the relative distance deviation using a pre-configured proportional-integral-derivative controller. The relative heading angle deviation is multiplied by the pre-acquired chassis wheelbase parameters to generate a lateral moment offset value for correcting the heading. Based on the aforementioned longitudinal traction force value and the aforementioned lateral torque offset value, the chassis traction control signal is generated, which includes a left drive wheel torque command and a right drive wheel torque command.

[0012] Optionally, before the step of generating the basic longitudinal traction force value based on the relative distance deviation using a pre-configured proportional-integral-derivative controller, the method further includes: Obtain the current ground adhesion coefficient of the environment in which the docking target area is located; Compare the current ground adhesion coefficient with a preset standard adhesion coefficient threshold; When the current ground adhesion coefficient is less than the standard adhesion coefficient threshold, the integral gain parameter in the pre-configured proportional-integral-derivative regulator is adjusted downwards. The derivative gain parameter in the pre-configured proportional-integral-derivative (PID) controller is adjusted upwards to obtain an updated PID controller. The step of inputting the relative distance deviation into the updated proportional-integral-derivative (PID) controller to trigger the generation of a basic longitudinal traction force value based on the relative distance deviation according to the updated PID controller.

[0013] Optionally, after the step of causing the chassis actuator to perform a matching docking control action according to the chassis traction control signal, the method further includes: During the process of the chassis actuator performing the docking control action, the actual docking trajectory of the unmanned vehicle is extracted in real time; Obtain the pre-constructed theoretical docking trajectory; Calculate the dynamic following error value between the actual docking trajectory and the theoretical docking trajectory; Determine whether the dynamic following error value is greater than a preset safety tolerance threshold; When the dynamic following error value is greater than the safety tolerance threshold, an emergency braking command is generated; The emergency braking command is sent to the chassis actuator; The chassis actuator is controlled to stop performing the docking control action and enter the parking lock state.

[0014] Optionally, the step of calculating the dynamic following error value between the actual docking trajectory and the theoretical docking trajectory includes: Within a preset two-dimensional spatial plane, the actual docking trajectory is discretized into a sequence of actual trajectory points; The theoretical docking trajectory is discretized into a sequence of theoretical trajectory points; Iterate through each actual trajectory point in the sequence of actual trajectory points; Calculate the Euclidean distance between the actual trajectory point and the nearest theoretical trajectory point in the sequence of theoretical trajectory points to obtain the set of local distance deviations; The maximum Euclidean distance value extracted from the set of local distance deviations is the dynamic following error value.

[0015] A high-precision relative pose estimation and docking device for unmanned vehicles based on monocular vision, the device comprising: The image acquisition module is used to acquire monocular vision images containing the docking target area collected by the monocular vision perception device installed at the front of the unmanned vehicle; The edge feature extraction module is used to perform multi-level edge feature extraction processing on the monocular vision image to obtain a multi-scale edge pixel cluster of the docking target region. The texture feature extraction module is used to perform multi-level texture feature extraction processing on the monocular vision image to obtain local texture gradient features. The three-dimensional spatial reconstruction module is used to perform three-dimensional spatial reconstruction processing on the multi-scale edge pixel cluster to obtain an initial three-dimensional point cloud structure. The pose resolution module is used to perform pose resolution processing on the initial three-dimensional point cloud structure using the local texture gradient features to obtain an initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system. The curvature feature extraction module is used to extract motion trajectory curvature features from the pose state data of the unmanned vehicle in continuous frames within the historical sampling period; The pose compensation and control module is used to perform spatiotemporal kinematic compensation processing on the initial relative pose reference sequence using the curvature features of the motion trajectory, generate the target relative pose result and convert it into a chassis traction control signal, so that the chassis actuator can perform matching docking control actions according to the chassis traction control signal.

[0016] An electronic device, the electronic device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to execute the above-described high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision.

[0017] A computer-readable storage medium storing computer instructions for causing a computer to execute the above-described high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision.

[0018] The technical advantages of the technical solution provided in this application are: This application presents a high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision. Addressing the shortcomings of traditional docking schemes based on static image deviations, which are susceptible to interference and struggle to construct accurate spatial references, this method extracts multi-level edge and texture features from the monocular vision image. This is followed by 3D spatial reconstruction and the generation of an initial relative pose reference sequence within a pre-defined virtual 3D coordinate system. This solves the problem of traditional schemes being susceptible to interference from local noise and edge blurring due to the reliance on single 2D visual features. Compared to traditional methods that rely solely on the geometric center of the calibration plate for control, this application elevates multi-scale edge and local texture gradient features to a 3D point cloud structure for comprehensive pose analysis. This significantly enriches the 3D spatial representation of the docking target area and enhances its resistance to interference under complex lighting and rapid viewpoint changes, resulting in a substantial improvement in the robustness of the initial spatial pose estimation.

[0019] Furthermore, addressing the shortcomings of traditional single-frame static control schemes that neglect dynamic trajectory changes in the vehicle, leading to significant errors at the docking end, this application extracts the curvature features of the chassis's motion trajectory across consecutive frames within a historical sampling period. This feature is then used to perform spatiotemporal kinematic compensation on the initial pose reference sequence, ultimately generating a control signal that converts into chassis traction force. This solves the problem of smooth adjustment when the vehicle is approaching with high curvature or non-linear maneuvers. Compared to traditional direct control commands that ignore dynamic kinematic changes, this application introduces the real-time motion trajectory curvature of the autonomous vehicle as a compensation variable into the pose solution loop, achieving adaptive matching in the spatiotemporal dimension. This allows the chassis actuators to flexibly fine-tune the traction force distribution based on the vehicle's instantaneous motion state, effectively improving the estimation accuracy of the relative distance and heading angle at the docking end, and significantly mitigating the collision risk caused by excessive chassis traction or dynamic lag. Attached Figure Description

[0020] Figure 1 This application provides an embodiment of a high-precision relative pose estimation and docking scenario for an unmanned vehicle based on monocular vision. Figure 2 This application provides an embodiment of a high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision. Figure 3 This application provides an embodiment of a high-precision relative pose estimation and docking device for unmanned vehicles based on monocular vision. Figure 4 An electronic device is described in an embodiment of this application; Figure 5 This is a computer-readable storage medium according to an embodiment of the present application. Detailed Implementation

[0021] like Figure 1 As shown, this is an embodiment of the present application illustrating a high-precision relative pose estimation and docking scenario for an unmanned vehicle based on monocular vision; Figure 2 As shown in the figure, this application provides an embodiment of a high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision. The method includes the following steps: Acquire monocular visual images containing the docking target area collected by a monocular visual perception device installed at the front of the unmanned vehicle; Perform multi-level edge feature extraction processing on the monocular vision image to obtain a multi-scale edge pixel cluster of the docking target region; Multi-level texture feature extraction processing is performed on the monocular vision image to obtain local texture gradient features; The multi-scale edge pixel cluster is subjected to three-dimensional spatial reconstruction processing to obtain an initial three-dimensional point cloud structure; The initial three-dimensional point cloud structure is subjected to pose analysis processing using the local texture gradient features to obtain an initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system. Extract motion trajectory curvature features from continuous frame pose state data of autonomous vehicles within a historical sampling period; The initial relative pose reference sequence is processed by spatiotemporal kinematic compensation using the curvature features of the motion trajectory to generate the target relative pose result and convert it into a chassis traction control signal, so that the chassis actuator can perform a matching docking control action according to the chassis traction control signal.

[0022] Optionally, the steps of performing multi-level edge feature extraction processing on the monocular vision image to obtain multi-scale edge pixel clusters of the docking target region, and performing multi-level texture feature extraction processing on the monocular vision image to obtain local texture gradient features include: The monocular visual image is mapped from the reference three primary color color space to the hue, saturation and brightness color space to obtain the converted image data; A pre-configured spatial smoothing mask is invoked to perform noise filtering on the transformed image data, resulting in smoothed image data; Extract the pixel brightness abrupt change regions from the smoothed image data; The pixel brightness abrupt change regions are aggregated to perform multi-level edge feature extraction processing, thereby obtaining a multi-scale edge pixel cluster of the docking target region; Extract the grayscale change rate between adjacent pixels in the smoothed image data to generate a basic texture gradient field; The smoothed image data is divided into multiple local spatial blocks of different scales; Within the multiple local spatial blocks of different scales, the direction and amplitude of the basic texture gradient field are statistically aggregated to perform multi-level texture feature extraction processing, thereby obtaining the local texture gradient features.

[0023] Preferably, before performing multi-level edge feature extraction and multi-level texture feature extraction on the monocular vision image, the image acquisition time, image width, image height, pixel arrangement order, and installation pose record of the monocular vision sensing device relative to the front end of the unmanned vehicle are read first. The image acquisition time is used as the basis for the same-frame association between the monocular vision image and the subsequently generated pixel brightness change region, multi-scale edge pixel cluster, basic texture gradient field, and local texture gradient features. The image width and image height are used as the boundary basis for subsequent pixel coordinate reading. The pixel arrangement order is used as the data reading basis for traversing each pixel in the monocular vision image according to row and column positions. The expected imaging orientation of the docking target area in the monocular vision image is determined according to the installation pose record. Before entering the color space mapping, the monocular vision image is not directly judged based on the gray center of the entire image. Instead, the image region to be refined is first determined from the monocular vision image according to the installation pose record and the expected imaging orientation. The image region to be refined is retained under the pixel coordinates corresponding to the monocular vision image, so that the subsequently generated pixel brightness change region, multi-scale edge pixel cluster, basic texture gradient field and local texture gradient features can all be referenced back to the same monocular vision image, the same image acquisition time and the same pixel coordinates, avoiding the separation of edge refinement results and texture refinement results in spatial reference.

[0024] Preferably, the pre-configured spatial smoothing mask is not a fixed application of a standard mean filter window, but is pre-configured based on the imaging characteristics of the docking target area in a near-field docking scenario for unmanned vehicles. Specifically, a spatial smoothing mask configuration record is first established, including a mask coverage record, a center pixel weight record, a neighboring pixel weight record, an edge preservation determination record, and a scale switching record. The mask coverage record limits the range of adjacent pixels covered by the pre-configured spatial smoothing mask each time it participates in noise filtering. The range of adjacent pixels consists of the positions of pixels around the currently processed pixel within the image area to be refined that participate in weighted processing. The center pixel weight record limits the retention ratio of the currently processed pixel in noise filtering. The neighboring pixel weight record limits the participation ratio of each adjacent pixel within the range of adjacent pixels in noise filtering. The edge preservation determination record limits the avoidance of excessive smoothing of the pixel brightness change area when there are significant brightness jumps between adjacent pixels. The scale switching record adjusts the range of adjacent pixels of the pre-configured spatial smoothing mask as the docking target area gradually transitions from far-field imaging to near-field imaging. In the above manner, the spatial smoothing mask configuration record can provide an executable parameter source for the pre-configured spatial smoothing mask before noise filtering, so that the converted image data retains the boundary contour changes of the docking target area while removing isolated noise areas, and the boundary contour changes continue to participate in the subsequent extraction of pixel brightness change areas.

[0025] Preferably, when mapping the monocular visual image from a reference three-primary-color space to a hue-saturation-luminance color space, the three-primary-color channel components corresponding to each pixel in the monocular visual image are read respectively, and the corresponding hue component, saturation component, and luminance component are generated according to the relative magnitude relationship between the three-primary-color channel components, thereby obtaining the converted image data. The luminance component in the converted image data continues to serve as the direct data source for extracting the brightness abrupt change region of the pixel. The hue component and saturation component in the converted image data are used to distinguish the color distribution difference between the surface mark of the docking target area and the environmental background. When the docking target area is affected by shadows, reflections, or partial occlusion, the converted image data no longer mixes color changes and brightness changes into a single grayscale difference. Instead, the luminance component in the converted image data participates in edge refinement, and the hue component and saturation component in the converted image data participate in edge region stability verification. Based on the result of the edge region stability verification, interfering pixels that are discontinuous with the surface mark of the docking target area are screened out, thereby ensuring that subsequent noise filtering and edge aggregation processing have the same image data basis.

[0026] Preferably, when the pre-configured spatial smoothing mask is invoked to perform noise filtering on the converted image data, the range of adjacent pixels matching the image region to be refined is first determined according to the spatial smoothing mask configuration record. Then, according to the center pixel weight record and the neighboring pixel weight record, the luminance component in the converted image data is subjected to neighborhood weighting processing to obtain initial smoothed luminance data. Subsequently, the initial smoothed luminance data is compared with the luminance component in the converted image data to obtain a difference comparison result. When the difference comparison result indicates that the currently processed pixel is located near the boundary of the docking target region, the smoothing effect of each adjacent pixel within the adjacent pixel range on the currently processed pixel is reduced according to the edge preservation determination record. When the difference comparison result indicates that the currently processed pixel is located in the background texture region or isolated noise region, the smoothing effect of each adjacent pixel within the adjacent pixel range on the currently processed pixel is enhanced according to the neighboring pixel weight record. After the above noise filtering process, the initial smoothed brightness data and the hue and saturation components in the converted image data together form the smoothed image data. The smoothed image data continues to serve as the common input data for extracting the pixel brightness abrupt change regions and generating the basic texture gradient field. The noise filtering results of the background texture region and the isolated noise region are still reflected in the brightness component of the smoothed image data.

[0027] Preferably, when extracting the pixel brightness abrupt change region from the smoothed image data, the brightness difference between each pixel and its neighboring pixels is first read from the brightness component of the smoothed image data. The brightness differences between horizontally adjacent pixels, vertically adjacent pixels, and diagonally adjacent pixels are recorded as brightness change direction data. Subsequently, the brightness change intensity around each pixel is calculated based on the brightness change direction data, and the brightness change intensity is compared with the corresponding neighborhood background change intensity. Pixels with a brightness change intensity higher than the corresponding neighborhood background change intensity are classified as candidate brightness abrupt change pixels. The candidate brightness abrupt change pixels are further screened for consistency by combining the hue and saturation components in the converted image data. When a candidate brightness abrupt change pixel has a boundary change in the brightness component of the smoothed image data and maintains a continuous correspondence with the surface marker of the docking target area in the hue or saturation component of the converted image data, the candidate brightness abrupt change pixel is retained as the pixel brightness abrupt change region. The resulting pixel brightness abrupt change region is not an isolated edge point in a single frame image, but a brightness abrupt change region associated with the surface contour and marker distribution of the docking target region. The brightness abrupt change region continues to serve as the pixel source for subsequent edge aggregation processing.

[0028] Preferably, when aggregating the pixel brightness abrupt change regions for multi-level edge feature extraction, firstly, based on the spatial connectivity of adjacent pixels in the pixel brightness abrupt change regions, the candidate brightness abrupt change pixels that are connected to each other are organized into edge connected segments. Then, based on the extension direction, length range, and local bending state of the edge connected segments, short isolated segments formed by dust, reflective spots, or image compression noise are removed. The edge connected segments after the removal process are further aggregated according to different scales: at a smaller scale, the narrow edges in the local contour of the docking target region are retained; at a larger scale, multiple edge connected segments with consistent direction and continuous spacing are merged into contour edge groups; at a scale corresponding to the geometric boundary of the docking target region, the contour edge groups are integrated into the multi-scale edge pixel cluster that can express the docking opening, docking plane boundary, or target identifier boundary. The multi-scale edge pixel cluster then participates in the three-dimensional space reconstruction process. Therefore, during the formation process, the multi-scale edge pixel cluster retains the pixel coordinates, scale source, and edge direction corresponding to the monocular vision image, so that the multi-scale edge pixel cluster can provide a traceable two-dimensional edge basis for the subsequent three-dimensional space reconstruction process. The two-dimensional edge basis continues to serve as the basis for the reconstruction of the initial three-dimensional point cloud structure.

[0029] Preferably, when extracting the grayscale change rate between adjacent pixels in the smoothed image data, grayscale representation data is first formed from the luminance component of the smoothed image data. Then, the grayscale change amounts between horizontally adjacent pixels, vertically adjacent pixels, and diagonally adjacent pixels in the grayscale representation data are read respectively. Based on the pixel interval of adjacent pixels in the image, the grayscale change amounts are converted into grayscale change rates. The grayscale change rates are then written into a basic texture gradient field according to pixel position and change direction. Each position in the basic texture gradient field corresponds to a pixel in the smoothed image data, and each position records the grayscale change rate and grayscale change amplitude in multiple directions around the pixel. The basic texture gradient field has a different function than the pixel brightness abrupt change region. The pixel brightness abrupt change region is used to express the boundary abrupt change of the docking target region, while the basic texture gradient field is used to express the directional distribution of the surface texture, marker texture, and boundary adjacent texture inside the docking target region. Both are generated on the same smoothed image data and subsequently serve the multi-scale edge pixel cluster and the local texture gradient features, respectively.

[0030] Preferably, when dividing the smoothed image data into multiple local spatial blocks of different scales, firstly, based on the imaging range of the docking target region in the smoothed image data, a block division range covering the docking target region is established. Then, within the block division range, multiple local spatial blocks of different scales are generated according to different pixel coverage scales. The smaller-scale local spatial blocks are used to preserve the grayscale changes near the corners, marker textures, and narrow boundaries of the docking target region, while the larger-scale local spatial blocks are used to express the overall texture direction and light-dark gradient trend of the surface of the docking target region. When the unmanned vehicle approaches the docking target region, the multiple local spatial blocks of different scales can cover the locally magnified area caused by the change in viewing angle, so that the basic texture gradient field can still be organized into a unified block structure at different imaging distances. After the formation of the multiple local spatial blocks of different scales, each local spatial block retains its pixel coordinate range in the smoothed image data so that the direction and amplitude statistics of the basic texture gradient field can be backfilled into the corresponding image position. The image position continues to correspond to the pixel coordinates of the monocular vision image, so that the multiple local spatial blocks of different scales can maintain spatial reference consistency with the multi-scale edge pixel cluster.

[0031] Preferably, when statistically aggregating the direction and amplitude of the basic texture gradient field within the multiple local spatial blocks of different scales, the grayscale change rate within the coverage area of ​​each local spatial block is first read, and a direction distribution record is formed based on the direction of the grayscale change rate. The direction distribution record is then superimposed with the grayscale change amplitude in the corresponding direction to form a direction amplitude distribution record. For the same local spatial block, if the direction amplitude distribution records in multiple directions are scattered, the local spatial block is marked as a texture disturbance block; if the direction amplitude distribution records are concentrated within the direction range corresponding to the boundary direction of the docking target area or the surface identification direction, the local spatial block is marked as a texture stable block. The texture perturbation block and the texture stabilization block continue to participate in the generation of the local texture gradient features. The texture perturbation block is used to reduce the influence of local noise or reflective areas on texture expression, and the texture stabilization block is used to enhance the texture direction expression consistent with the structural boundary of the docking target region. Both the texture perturbation block and the texture stabilization block are retained within the pixel coordinate range corresponding to the local spatial block, so as to establish a spatial correspondence with the initial three-dimensional point cloud structure in the future.

[0032] Preferably, when generating the local texture gradient features, the directional amplitude distribution records at each scale are arranged according to the pixel coordinate range of the local spatial blocks in the smoothed image data, and the directional amplitude distribution records of the same image position at different scales are associated to form a cross-scale texture association record. The cross-scale texture association record does not replace the basic texture gradient field, but rather organizes the dispersed direction and amplitude information in the basic texture gradient field into blocks, so that the local texture gradient features can simultaneously express the local texture direction, local texture intensity, and the stability of the texture direction at different scales. The local texture gradient features are then used to perform pose resolution processing on the initial 3D point cloud structure. Therefore, when the local texture gradient features are generated, the pixel coordinate range of the corresponding local spatial block is retained, and the pixel coordinate range of the local spatial block continues to correspond to the pixel coordinates of the monocular vision image, so that the local texture gradient features can establish a spatial correspondence with the initial 3D point cloud structure obtained from the multi-scale edge pixel clusters.

[0033] Preferably, in a scenario where the unmanned vehicle slowly approaches a charging base station or docking station, the docking target area typically includes a border, a guide opening, planar markings, and a local reflective area. The pre-configured spatial smoothing mask first performs noise filtering on isolated bright spots and low-amplitude random textures in the converted image data. Both isolated bright spots and low-amplitude random textures serve as the image representation of isolated noise areas in the converted image data, and the noise filtering process forms the smoothed image data. The brightness component in the smoothed image data is then used to extract the pixel brightness abrupt change regions. These pixel brightness abrupt change regions are then aggregated to form the multi-scale edge pixel cluster, which is used to represent the external boundary and guide opening boundary of the docking target area. Simultaneously, the grayscale change rate in the smoothed image data is organized into a basic texture gradient field. This basic texture gradient field is statistically aggregated in terms of direction and amplitude within multiple local spatial blocks of different scales to form the local texture gradient features. These local texture gradient features are used to express the surface texture direction and marking texture distribution of the docking target area. The multi-scale edge pixel cluster and the local texture gradient feature provide data support from the perspectives of boundary geometry and texture direction, respectively. The multi-scale edge pixel cluster and the local texture gradient feature originate from the same smoothed image data and participate in subsequent three-dimensional space reconstruction and pose analysis processing under the pixel coordinates corresponding to the same monocular vision image.

[0034] Preferably, the above-mentioned multi-level edge feature extraction and multi-level texture feature extraction processes differ from the processing method that only extracts the center point of the two-dimensional calibration pattern. The processing does not compress the monocular visual image into a single center deviation. Instead, it first obtains the transformed image data through color space mapping, and then obtains the smoothed image data through the pre-configured spatial smoothing mask. On the smoothed image data, on the one hand, the pixel brightness abrupt change regions are extracted and aggregated to obtain the multi-scale edge pixel cluster; on the other hand, the basic texture gradient field is extracted and block-based statistically analyzed to obtain the local texture gradient features. The multi-scale edge pixel cluster continues to participate in the construction of the initial three-dimensional point cloud structure, and the local texture gradient features continue to participate in the pose analysis of the initial three-dimensional point cloud structure. This ensures that even in docking scenarios where illumination changes, viewing angle changes, and local texture interference coexist, the docking target area can still have its relative pose reference information jointly defined by the edge contour expressed by the multi-scale edge pixel cluster and the texture direction expressed by the local texture gradient features, thus providing a relatively stable data foundation for the subsequent generation of the initial relative pose reference sequence.

[0035] Optionally, the step of performing three-dimensional spatial reconstruction processing on the multi-scale edge pixel cluster to obtain an initial three-dimensional point cloud structure; and performing pose analysis processing on the initial three-dimensional point cloud structure using the local texture gradient features to obtain an initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system includes: Obtain the pre-calibrated set of internal calibration parameters of the monocular vision sensing device; The internal calibration parameter set is used to perform inverse projection calculation on the two-dimensional pixel coordinates in the multi-scale edge pixel cluster to generate a first three-dimensional spatial ray set; Calculate the spatial intersection points of the first three-dimensional spatial ray set within the preset depth constraint range; Spatial topology aggregation is performed on the spatial intersections to perform the three-dimensional spatial reconstruction process, resulting in the initial three-dimensional point cloud structure. The estimated normal vector of the docking target surface is determined based on the local texture gradient features. The initial 3D point cloud structure is projected onto the projection plane corresponding to the normal vector estimate to obtain the relative translation vector and the relative rotation feature set. The pose analysis process is performed based on the relative translation vector and the relative rotation feature set to obtain the initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system.

[0036] Preferably, the specific implementation process of obtaining the pre-calibrated internal calibration parameter set of the monocular vision perception device is as follows: After the monocular vision perception device is installed on the front end of the unmanned vehicle and the installation pose is fixed, multi-frame calibration image data containing a known geometric spacing calibration pattern is first collected, and calibration reference point pixel coordinate records are extracted from the multi-frame calibration image data; the calibration reference point pixel coordinate records are used to express the pixel coordinate distribution corresponding to the known geometric spacing calibration pattern at different imaging positions. Subsequently, based on the geometric positional correspondence between the pixel coordinate records of the calibration reference points and the known geometric spacing calibration pattern, the internal imaging parameter records of the monocular vision sensing device are obtained. These internal imaging parameter records include focal length parameter records, principal point position parameter records, pixel size conversion records, and distortion correction parameter records. The focal length parameter records describe the magnification relationship of spatial points on the image plane after imaging by the monocular vision sensing device. The principal point position parameter records describe the intersection position of the optical axis of the monocular vision sensing device and the image plane. The pixel size conversion records describe the conversion relationship between the pixel spacing and the geometric spacing of the image plane. The distortion correction parameter records describe the radial and tangential offsets generated in the lens edge region of the monocular vision sensing device. After undergoing repeated projection error verification, the imaging parameter records inside the monocular vision sensing device are written into the pre-calibrated internal calibration parameter set of the monocular vision sensing device. This allows the pre-calibrated internal calibration parameter set to provide a conversion basis between the two-dimensional pixel coordinates and the three-dimensional ray direction in the multi-scale edge pixel cluster during subsequent inverse projection calculations. The repeated projection error verification is used to remap the pixel coordinate records of the calibration reference point onto the image plane through the imaging parameter records inside the monocular vision sensing device, and to verify the positional difference between the remapped pixel coordinate distribution and the pixel coordinate records of the calibration reference point, so as to filter out the imaging parameter records inside the monocular vision sensing device that do not meet the calibration stability.

[0037] Preferably, the preset virtual three-dimensional coordinate system is configured before the unmanned vehicle performs the docking task. This configuration process does not rely on the image center point temporarily generated during docking, but is established based on the installation pose record of the monocular vision sensing device relative to the front end of the unmanned vehicle and the motion direction record of the unmanned vehicle chassis. The installation pose record originates from the installation pose calibration data of the monocular vision sensing device after installation, and this data expresses the installation height, orientation, and offset of the monocular vision sensing device relative to the front end of the unmanned vehicle. The motion direction record of the unmanned vehicle chassis originates from the forward direction record and the left-right wheel alignment record of the chassis during calibration driving, and this record serves as the motion reference for the chassis actuators in subsequent docking control actions. Specifically, the imaging optical center of the monocular vision sensing device is designated as the spatial reference origin. The forward direction of the unmanned vehicle towards the docking target area is configured as the spatial depth direction. The direction connecting the left and right wheels of the unmanned vehicle chassis is configured as the spatial lateral direction. The direction perpendicular to both the spatial depth direction and the spatial lateral direction is configured as the spatial height direction, thereby forming the preset virtual three-dimensional coordinate system. The preset virtual three-dimensional coordinate system is further associated with the pre-calibrated internal calibration parameter set of the monocular vision sensing device, so that the two-dimensional pixel coordinates in the multi-scale edge pixel cluster can be transformed into the preset virtual three-dimensional coordinate system after inverse projection calculation. The preset virtual three-dimensional coordinate system is also further associated with the motion direction of the chassis actuator, so that the subsequently obtained initial relative pose reference sequence can maintain the same spatial meaning as the forward and backward offset, left and right offset, and attitude deflection required for the docking control of the unmanned vehicle.

[0038] Preferably, before performing inverse projection calculations on the two-dimensional pixel coordinates in the multi-scale edge pixel cluster using the pre-calibrated internal calibration parameter set of the monocular vision perception device, the two-dimensional pixel coordinates, scale source, and edge direction corresponding to each edge pixel in the multi-scale edge pixel cluster are first read, and distortion correction processing is performed on the two-dimensional pixel coordinates. The distortion correction processing, based on the distortion correction parameter record, corrects the two-dimensional pixel coordinates near the edge region of the image plane to corrected two-dimensional pixel coordinates corresponding to the ideal pinhole imaging relationship. The corrected two-dimensional pixel coordinates are then combined with the principal point position parameter record, the focal length parameter record, and the pixel size conversion record for image plane normalization processing to obtain the imaging direction record corresponding to each edge pixel. Subsequently, the imaging direction record is converted into a three-dimensional ray direction extending outward from the spatial reference origin of the preset virtual three-dimensional coordinate system, and the three-dimensional ray direction is bound to the corresponding two-dimensional pixel coordinates, scale source, and edge direction to generate the first three-dimensional spatial ray set. Therefore, the first three-dimensional spatial ray set is not a simple combination of three-dimensional ray directions, but retains the position source, scale source, and edge direction of the edge pixels in the multi-scale edge pixel cluster in the monocular vision image, so that the subsequent spatial intersection calculation can continue to utilize the spatial coherence of the edge contour; the position source is expressed by the two-dimensional pixel coordinates, the scale source is expressed by the formation scale of the multi-scale edge pixel cluster in the multi-level edge feature extraction process, and the edge direction is expressed by the extension direction of the edge pixels in the multi-scale edge pixel cluster.

[0039] Preferably, when calculating the spatial intersection points of the first three-dimensional spatial ray set within a preset depth constraint range, the preset depth constraint range is pre-configured based on the allowable proximity distance between the docking target area and the unmanned vehicle, the installation pose record of the monocular vision sensing device at the front end of the unmanned vehicle, and the visible boundary size of the docking target area. The allowable proximity distance is used to limit the spatial depth search range of the unmanned vehicle during the docking process, the visible boundary size of the docking target area is used to limit the boundary range of the docking target area that can be imaged and observed in the preset virtual three-dimensional coordinate system, and the installation pose record is used to limit the spatial offset between the imaging optical center of the monocular vision sensing device and the front end of the unmanned vehicle. The preset depth constraint range is used to limit the effective value area of ​​the first three-dimensional spatial ray set in the spatial depth direction of the preset virtual three-dimensional coordinate system. Subsequently, for each of the three-dimensional ray directions in the first three-dimensional spatial ray set, multiple candidate spatial positions are generated in the spatial depth direction according to the preset depth constraint range. Based on the edge direction consistency, scale origin consistency, and spatial distance continuity between adjacent three-dimensional ray directions, spatial intersection points matching the contour of the docking target region are selected from the multiple candidate spatial positions. The edge direction consistency is used to determine whether the edge pixels corresponding to adjacent three-dimensional ray directions originate from similar contour extension directions. The scale origin consistency is used to determine whether the edge pixels corresponding to adjacent three-dimensional ray directions originate from the same or adjacent edge refinement scale. The spatial distance continuity is used to determine whether adjacent candidate spatial positions can form a continuous boundary under the preset virtual three-dimensional coordinate system. The spatial intersection points retain their source three-dimensional ray direction, source two-dimensional pixel coordinates, scale origin, and edge direction, enabling the spatial intersection points to express the spatial distribution of the border, guide opening, and target identifier boundary of the docking target region under the preset virtual three-dimensional coordinate system in subsequent spatial topology aggregation.

[0040] Preferably, when performing spatial topology aggregation on the spatial intersections for the 3D spatial reconstruction process, the spatial intersections are first organized into spatial edge segments based on their adjacent distances in the preset virtual 3D coordinate system, where the spatial intersections are continuously spaced and have similar edge directions. These spatial edge segments are then further aggregated according to their scale source, preserving spatial details at local corners at smaller scale sources and merging spatial edge segments with continuous directions at larger scale sources to form spatial contour segments. Subsequently, based on the relative positional relationship between the spatial contour segments, it is determined whether the spatial contour segments can collectively express the planar boundary, guide opening boundary, and target identifier boundary of the docking target area. Spatial contour segments that satisfy spatial coherence are then aggregated into the initial 3D point cloud structure. The spatial coherence relationship includes the continuity of endpoint distances, edge direction connection, and scale source connection between adjacent spatial contour segments. This spatial coherence relationship is used to exclude discontinuous spatial contour segments formed by isolated noise, reflective edges, or short missing edges. Each 3D point in the initial 3D point cloud structure retains its source spatial intersection, source 2D pixel coordinates, scale source, and edge direction. Therefore, the initial 3D point cloud structure can establish a correspondence with the multi-scale edge pixel cluster and continue to serve as the spatial basis for the local texture gradient features to participate in pose analysis processing.

[0041] Preferably, when determining the estimated normal vector of the docking target surface based on the local texture gradient features, the texture direction, texture intensity, and stability of the texture direction at different scales corresponding to each local spatial block in the local texture gradient features are first read, and the pixel coordinate range of the local spatial block is matched with the corresponding source two-dimensional pixel coordinates in the initial three-dimensional point cloud structure. Through the matching process, the local texture gradient features can be associated with the corresponding three-dimensional point positions in the initial three-dimensional point cloud structure; specifically, the matching process, based on the inclusion relationship between the pixel coordinate range of the local spatial block and the source two-dimensional pixel coordinates, groups the three-dimensional points corresponding to the source two-dimensional pixel coordinates located in the same local spatial block into the same texture-associated three-dimensional point range. Subsequently, within a range of texture-associated 3D points continuous with the surface texture direction of the docking target area, candidate surface directions are determined based on the spatial arrangement trend between the 3D points. The stability of the texture direction at different scales is used to filter these candidate surface directions, ensuring that candidate surface directions caused by reflective areas, partial occlusion areas, or isolated edge errors do not participate in the dominant direction determination. The reflective areas, partial occlusion areas, and isolated edge errors are all identified through the stability of the texture direction at different scales. When the corresponding texture direction changes dispersedly at different scales, the associated candidate surface directions have their weight reduced. The filtered candidate surface directions undergo direction consistency integration to obtain the normal vector estimate of the docking target surface. This direction consistency integration merges candidate surface directions with similar orientations within the texture-associated 3D point range and eliminates candidate surface directions that deviate significantly from the majority of candidate surface directions. The normal vector estimate expresses the orientation of the visible surface of the docking target area relative to the preset virtual 3D coordinate system and continues to serve as the orientation basis for projecting the initial 3D point cloud structure onto the corresponding projection plane.

[0042] Preferably, when projecting the initial 3D point cloud structure onto the projection plane corresponding to the normal vector estimate, a projection plane corresponding to the orientation of the docking target surface is first established based on the normal vector estimate and the distribution position of the spatial contour segments in the initial 3D point cloud structure. This projection plane is not an additional plane introduced independently of the initial 3D point cloud structure, but rather a geometric reference plane whose orientation is defined by the normal vector estimate and whose spatial position is defined by the initial 3D point cloud structure. This geometric reference plane continues to bear the projection result of the initial 3D point cloud structure. Subsequently, the 3D points in the initial 3D point cloud structure are mapped onto the projection plane along the vertical direction corresponding to the normal vector estimate, resulting in a projection point distribution record. This projection point distribution record continues to retain the source 3D points, source spatial intersections, source 2D pixel coordinates, and source local spatial blocks, enabling the projection point distribution record to simultaneously express the spatial position of the edge contour and the surface orientation of the texture direction. By recording the projection point distribution, the three-dimensional point distribution in the initial three-dimensional point cloud structure affected by the viewing angle tilt can be converted into a two-dimensional projection expression that is closer to the planar structure of the docking target surface, thereby providing a unified projection reference for subsequently obtaining the relative translation vector and the relative rotation feature set; the two-dimensional projection expression continues to remain in the projection plane, and is expressed by the distribution position of the projection point distribution record in the projection plane.

[0043] Preferably, when obtaining the relative translation vector, the projection center position that can express the center position of the outline of the docking target area is first extracted from the projection point distribution record, and the position difference between the projection center position and the spatial reference origin of the preset virtual three-dimensional coordinate system is calculated to obtain the relative translation vector. The projection center position originates from the projection point positions in the projection point distribution record that can jointly express the planar boundary, guide opening boundary, and target mark boundary of the docking target area, and the projection center position continues to retain its projection coordinates in the projection plane. The relative translation vector includes a distance offset along the spatial depth direction, a lateral offset along the spatial lateral direction, and a height offset along the spatial height direction. The distance offset along the spatial depth direction is used to express the approach distance between the unmanned vehicle and the docking target area, the lateral offset along the spatial lateral direction is used to express the left-right offset between the unmanned vehicle and the docking target area, and the height offset along the spatial height direction is used to express the vertical offset of the docking target area in the field of view of the monocular vision perception device. The relative translation vector continues to participate in the pose analysis process and provides a data source for the translation state in the subsequent initial relative pose reference sequence; the translation state is jointly expressed by the distance offset in the spatial depth direction, the lateral offset in the spatial lateral direction, and the height offset in the spatial height direction.

[0044] Preferably, when obtaining the relative rotation feature set, the boundary direction record of the docking target area is first determined based on the main extension direction of the spatial contour segment in the projection plane according to the projection point distribution record. The boundary direction record of the docking target area is then subjected to orientation deviation analysis with the spatial lateral direction and spatial height direction in the preset virtual three-dimensional coordinate system to obtain the in-plane deflection record. Subsequently, based on the orientation difference between the normal vector estimate and the spatial depth direction in the preset virtual three-dimensional coordinate system, the surface orientation deflection record is obtained. The in-plane deflection record is used to express the rotation state of the docking target area in the projection plane, and the surface orientation deflection record is used to express the pitch deflection and yaw deflection of the docking target area relative to the forward direction of the unmanned vehicle. The in-plane deflection record and the surface orientation deflection record together form the relative rotation feature set, which continues to participate in the pose analysis processing and provides a data source for the attitude state in the subsequent initial relative pose reference sequence. The attitude state is jointly expressed by the in-plane deflection record and the surface orientation deflection record.

[0045] Preferably, when performing pose analysis based on the relative translation vector and the relative rotation feature set, the relative translation vector and the relative rotation feature set obtained at the same image acquisition time are first time-bound to form a single-frame relative pose record; the image acquisition time is derived from the acquisition timestamp of the monocular vision image, and the image acquisition time is used to maintain the same-frame correspondence between the relative translation vector, the relative rotation feature set, the multi-scale edge pixel cluster, and the local texture gradient features. The single-frame relative pose record includes the distance offset in the spatial depth direction, the lateral offset in the spatial lateral direction, the height offset in the spatial height direction, the in-plane deflection record, and the surface orientation deflection record. Subsequently, the multiple single-frame relative pose records formed during continuous image acquisition are arranged in chronological order. Abnormal single-frame relative pose records caused by momentary occlusion or reflection are removed based on the displacement continuity and attitude continuity between the single-frame relative pose records. The displacement continuity is used to determine whether the spatial depth offset, spatial lateral offset, and spatial height offset in adjacent image acquisition times conform to the spatial change trend when the unmanned vehicle approaches the docking target area. The attitude continuity is used to determine whether the in-plane deflection records and surface orientation deflection records in adjacent image acquisition times conform to the attitude change trend when the unmanned vehicle approaches the docking target area. The single-frame relative pose records after the removal process are arranged in chronological order to form an initial relative pose reference sequence located in the preset virtual three-dimensional coordinate system. The initial relative pose reference sequence continues to serve as input for spatiotemporal kinematic compensation processing, enabling the subsequent motion trajectory curvature features to compensate for the initial relative pose reference sequence within the same preset virtual three-dimensional coordinate system.

[0046] Preferably, in a scenario where the unmanned vehicle approaches the docking opening of the charging base station, the multi-scale edge pixel cluster typically includes two-dimensional pixel coordinates corresponding to the charging base station border, the guide opening boundary, and the target identifier boundary. The pre-calibrated monocular vision perception device's internal calibration parameter set first converts the two-dimensional pixel coordinates into the first three-dimensional spatial ray set. The first three-dimensional spatial ray set then combines with the preset depth constraint range to form the spatial intersection point. After spatial topological aggregation, the spatial intersection point forms the initial three-dimensional point cloud structure that can express the spatial relationship between the charging base station border, the guide opening boundary, and the target identifier boundary. Simultaneously, the portion of the local texture gradient feature that is continuous with the charging base station planar identifier and border texture direction is used to determine the normal vector estimate. The normal vector estimate further defines the projection surface, enabling the initial three-dimensional point cloud structure to be projected as the projection point distribution record. The projection point distribution record is analyzed by translation and rotation to obtain the relative translation vector and the relative rotation feature set, thereby forming the initial relative pose reference sequence. The initial relative pose reference sequence does not only depend on the two-dimensional center pixel deviation, but also comes from the three-dimensional spatial distribution of the edge contour, the texture direction of the docking target surface, and the spatial reference relationship under the preset virtual three-dimensional coordinate system.

[0047] Optionally, the step of extracting motion trajectory curvature features from the pose state data of the unmanned vehicle in continuous frames within the historical sampling period includes: The inertial measurement unit inside the unmanned vehicle records the acceleration time series and angular velocity time series during the historical sampling period. The continuous frame pose state data includes the acceleration time series and angular velocity time series. Perform a first time accumulation operation on the acceleration time series to obtain the velocity time series; Perform a second time accumulation operation on the velocity time series to obtain the displacement time series; Based on the displacement time series and the angular velocity time series, the continuous frame pose state data is generated; Determine the pose points of three adjacent frames in the continuous frame pose state data; Calculate the spatial dimensions of the circumcircle of the spatial triangle formed by the pose points of the three adjacent frames; The curvature characteristics of the motion trajectory are calculated based on the circumscribed circle space dimensions.

[0048] Preferably, before reading the acceleration time series and angular velocity time series recorded by the inertial measurement unit inside the unmanned vehicle within the historical sampling period, the historical sampling period is configured based on the sampling frequency when the unmanned vehicle performs docking control actions, the image acquisition time of the monocular vision image, and the control refresh cycle of the chassis actuator. The sampling frequency is used to limit the number of times the inertial measurement unit generates acceleration and angular velocity sampling data per unit time. The image acquisition time is used to limit the visual pose resolution time corresponding to the monocular vision image. The control refresh cycle is used to limit the time interval between the chassis actuator receiving updated control quantities. The historical sampling period is jointly determined based on the sampling frequency, the image acquisition time, and the control refresh cycle. The historical sampling period is used to limit the time range of inertial data participating in the extraction of motion trajectory curvature features, so that the acceleration time series and the angular velocity time series can cover the continuous motion process of the unmanned vehicle approaching the docking target area. After configuring the historical sampling period, the acceleration sampling data and angular velocity sampling data output by the inertial measurement unit are read in the order of sampling timestamps within the historical sampling period. The acceleration sampling data is arranged into an acceleration time series according to the sampling timestamps, and the angular velocity sampling data is arranged into an angular velocity time series according to the same sampling timestamps. Both the acceleration time series and the angular velocity time series retain the corresponding sampling timestamps, so that the subsequent first time accumulation calculation, second time accumulation calculation, and continuous frame pose state data generation processing can be performed under the same time reference, avoiding the mixing of acceleration changes and angular velocity changes at different times into the same frame pose state.

[0049] Preferably, before participating in the first time accumulation calculation, the acceleration time series is first corrected by combining the acceleration zero-bias record generated by the inertial measurement unit in a static calibration state, and then subjected to gravity direction subtraction processing by combining the gravity direction record generated by the unmanned vehicle in a horizontally stationary state, so as to obtain a corrected acceleration time series that conforms to the motion direction of the unmanned vehicle. The acceleration zero-bias record is used to express the acceleration reference offset still generated by the inertial measurement unit when there is no motion input. The acceleration zero-bias record is obtained by the inertial measurement unit through continuous sampling in the static calibration state and continues to participate in the acceleration zero-bias correction. The gravity direction record is used to express the directional distribution of gravitational acceleration under the coordinate reference of the inertial measurement unit itself. The gravity direction record is obtained by the acceleration sampling data of the unmanned vehicle in the horizontally stationary state and continues to participate in the gravity direction subtraction processing. The corrected acceleration time series continues to use the sampling timestamp of the acceleration time series, so that the corrected acceleration time series can maintain a time correspondence with the angular velocity time series. The corrected acceleration time series does not change the technical meaning of the acceleration time series as the basic data for the first time accumulation calculation. Instead, it removes the static offset and gravity component before the first time accumulation calculation, so that the velocity time series obtained from the corrected acceleration time series is closer to the chassis motion state of the unmanned vehicle in the historical sampling period.

[0050] Preferably, when performing the first time accumulation operation on the acceleration time series to obtain the velocity time series, the acceleration sampling data of adjacent sampling moments in the corrected acceleration time series are first read according to the chronological order of the sampling timestamps. Based on the sampling time interval between adjacent sampling moments, each acceleration change is converted into a velocity increment record within the corresponding sampling time interval. The velocity increment records are then sequentially superimposed onto the initial velocity record of the unmanned vehicle at the start of the historical sampling period according to the sampling timestamps to form the velocity time series. The initial velocity record is provided by the chassis velocimeter output before the unmanned vehicle enters the historical sampling period, or configured as a zero-speed state when the unmanned vehicle is in a stationary start-up docking state. The initial velocity record continues to serve as the starting velocity basis for accumulating the velocity increment records. Each velocity state in the velocity time series corresponds to a sampling timestamp in the corrected acceleration time series. The velocity time series continues to serve as input data for the second time accumulation operation, enabling subsequent displacement time series to be generated in a time sequence consistent with the actual movement process of the unmanned vehicle.

[0051] Preferably, when performing a second time accumulation operation on the velocity time series to obtain the displacement time series, firstly, based on the velocity state at adjacent sampling moments in the velocity time series and the corresponding sampling time interval, a displacement increment record is calculated for each sampling time interval. The displacement increment record is used to express the position change of the unmanned vehicle along its own direction of motion within the corresponding sampling time interval. Subsequently, the displacement increment records are sequentially superimposed onto the initial position record of the unmanned vehicle at the start of the historical sampling period according to the sampling timestamp, forming the displacement time series. The initial position record is used to express the spatial position of the unmanned vehicle relative to the preset virtual three-dimensional coordinate system at the start of the historical sampling period. The initial position record continues to serve as the starting position basis for accumulating the displacement increment records, and the displacement time series continues to retain the sampling timestamp of the velocity time series. Through the second time accumulation operation, the velocity state in the velocity time series is converted into the position state in the displacement time series, and the displacement time series can provide a position source for subsequently generating the continuous frame pose state data.

[0052] Preferably, before participating in the generation of the continuous frame pose state data, the angular velocity time series is first corrected by combining the angular velocity zero-bias record formed by the inertial measurement unit in the static calibration state to obtain the corrected angular velocity time series. The angular velocity zero-bias record is used to express the angular velocity reference offset still generated by the inertial measurement unit when there is no rotational input. The angular velocity zero-bias record is obtained by the inertial measurement unit through continuous sampling in the static calibration state and continues to participate in the angular velocity zero-bias correction. The corrected angular velocity time series continues to retain the sampling timestamp of the angular velocity time series. Subsequently, the corrected angular velocity time series is subjected to angle accumulation processing according to the order of the sampling timestamps to obtain the attitude change time series. The attitude change time series is used to express the heading, pitch, and roll changes of the unmanned vehicle in the historical sampling period. The attitude change time series and the displacement time series are bound using the same sampling timestamp, so that the position state in the displacement time series and the attitude state in the attitude change time series can jointly participate in the generation of the continuous frame pose state data.

[0053] Preferably, when generating the continuous frame pose state data based on the displacement time series and the angular velocity time series, each position state in the displacement time series is first paired with the attitude state in the attitude change time series under the same sampling timestamp to form a single frame pose state record. The single frame pose state record includes the position state, attitude state, velocity state, and sampling timestamp of the unmanned vehicle at the corresponding sampling timestamp; wherein the position state originates from the displacement time series, the attitude state originates from the attitude change time series, the velocity state originates from the velocity time series, and the sampling timestamp originates from the same sampling moment retained by the acceleration time series and the angular velocity time series. Multiple single frame pose state records are arranged in chronological order according to the sampling timestamps to form the continuous frame pose state data; the continuous frame pose state data can therefore simultaneously express the continuous position and attitude changes of the unmanned vehicle within the historical sampling period, providing a continuous pose basis for determining the pose points of subsequent adjacent three frames.

[0054] Preferably, when determining the three adjacent frame pose points in the continuous frame pose state data, the single frame pose state records in the continuous frame pose state data are first traversed according to the sampling timestamp, and the corresponding position states are read from three temporally consecutive single frame pose state records respectively. The three position states are then used as the three adjacent frame pose points in sequence. Each pose point in the three adjacent frame pose points includes the corresponding spatial position and the sampling timestamp. The spatial position originates from the position state in the displacement time sequence, and the sampling timestamp originates from the single frame pose state record in the continuous frame pose state data. After generation, the three adjacent frame pose points are further associated with the attitude states in the corresponding three single frame pose state records, so that the three adjacent frame pose points not only express the curvature trend of the unmanned vehicle trajectory at the spatial position, but also reflect the angular velocity change state of the unmanned vehicle at the same sampling timestamp. Therefore, when calculating the circumscribed circle spatial dimensions, the calculation will not deviate from the temporal order and attitude change source of the continuous frame pose state data.

[0055] Preferably, before calculating the circumcircle size of the spatial triangle formed by the pose points of the three adjacent frames, the geometric unfolding degree of the spatial triangle is first determined based on the point distance relationship between the pose points of the three adjacent frames. The point distance relationship originates from the spatial position difference between any two pose points in the three adjacent frames, and the geometric unfolding degree is used to express the unfolding state of the spatial triangle relative to the straight trajectory. When the point distance relationship indicates that the unmanned vehicle is close to linear motion within the continuous sampling time, the spatial triangle is marked as a low-curvature trajectory segment, and the low-curvature trajectory segment continues to participate in the calculation of the circumcircle size to obtain a larger circumcircle size. When the point distance relationship indicates that the unmanned vehicle has turning motion within the continuous sampling time, the spatial triangle is marked as a turning trajectory segment, and the turning trajectory segment continues to participate in the calculation of the circumcircle size to obtain the circumcircle size corresponding to the turning radius. The above processing ensures that the spatial dimension of the circumcircle of the spatial triangle and the curvature of the unmanned vehicle's trajectory have the same physical meaning; that is, the larger the spatial dimension of the circumcircle, the lower the curvature of the trajectory; and the smaller the spatial dimension of the circumcircle, the higher the curvature of the trajectory.

[0056] Preferably, when calculating the circumcircle space size of the spatial triangle formed by the pose points of the three adjacent frames, three side length records are first obtained based on the spatial position difference between any two pose points in the three adjacent frames. These three side length records are used to express the span of the local motion trajectory formed by the pose points of the three adjacent frames within the historical sampling period. Subsequently, the geometric unfolding degree of the spatial triangle is determined based on the three side length records, and the circumcircle space size that can simultaneously pass through the pose points of the three adjacent frames is calculated based on the three side length records and the geometric unfolding degree. The circumcircle space size is used to express the local trajectory curvature radius corresponding to the pose points of the three adjacent frames. The circumcircle space size continues to retain the corresponding intermediate sampling timestamp, which is derived from the pose point located at the middle position in time among the pose points of the three adjacent frames, so that the motion trajectory curvature feature calculated subsequently can establish a temporal correlation with the corresponding motion moment in the continuous frame pose state data.

[0057] Preferably, when calculating the curvature feature of the motion trajectory based on the circumcircle spatial dimensions, each circumcircle spatial dimension is first arranged according to the corresponding intermediate sampling timestamp to form a curvature radius time record. The curvature radius time record is used to express the bending radius change of different local trajectory segments of the unmanned vehicle within the historical sampling period. Subsequently, according to the reciprocal relationship between the circumcircle spatial dimensions and the trajectory curvature, the curvature radius time record is converted into a curvature time record. The curvature value in the curvature time record is used to express the degree of curvature of the corresponding local trajectory segment. The curvature time record is further combined with the velocity state in the velocity time series for velocity consistency verification. When the velocity state corresponding to a certain curvature value is not continuous with the velocity state change in adjacent sampling timestamps, the curvature value is marked as an abnormal curvature value, and the abnormal curvature value is smoothed and corrected according to the time continuity relationship of adjacent curvature values ​​to obtain the motion trajectory curvature feature. The motion trajectory curvature feature is therefore derived from both the circumcircle spatial dimensions and the velocity time series, and can express the local steering intensity and motion continuity of the unmanned vehicle within the historical sampling period.

[0058] Preferably, the motion trajectory curvature feature, after being formed, is not used as an isolated value, but continues to establish a correspondence with the sampling timestamp, position state, attitude state, and velocity state in the continuous frame pose state data to form a time-referenced motion trajectory curvature feature. The time-referenced motion trajectory curvature feature includes the curvature value, curvature change trend, associated position state, and associated attitude state corresponding to the sampling timestamp; wherein, the curvature value originates from the curvature time record, the curvature change trend originates from the curvature value change under adjacent sampling timestamps, the associated position state originates from the displacement time series, and the associated attitude state originates from the attitude change time series. The time-referenced motion trajectory curvature feature continues to participate in subsequent spatiotemporal kinematic compensation processing, enabling the initial relative pose reference sequence to read the trajectory curvature state adjacent to the current docking time during compensation, rather than adjusting solely based on the static pose result of a single frame of the monocular visual image.

[0059] Preferably, in one scenario, when the unmanned vehicle approaches the docking opening of the charging base station, if the unmanned vehicle travels in an approximately straight line, the spatial triangle formed by the pose points of three adjacent frames in the continuous frame pose state data corresponds to the low-curvature trajectory segment. The circumscribed circle space size corresponding to the low-curvature trajectory segment is relatively large, resulting in a lower curvature value in the obtained motion trajectory curvature feature. If the unmanned vehicle performs a large-curvature turn to avoid obstacles or correct lateral offset, the spatial triangle formed by the pose points of three adjacent frames corresponds to the turning trajectory segment. The geometric unfolding degree corresponding to the turning trajectory segment is relatively high, the circumscribed circle space size corresponding to the turning trajectory segment is reduced, resulting in an increased curvature value in the obtained motion trajectory curvature feature. In this way, the motion trajectory curvature feature can transform the non-linear approach state of the unmanned vehicle within the historical sampling period into a data expression that can participate in spatiotemporal kinematic compensation processing. This allows the subsequent spatiotemporal kinematic compensation processing of the initial relative pose reference sequence using the motion trajectory curvature feature to adjust the pose reference of the docking end according to the trajectory curvature degree of the unmanned vehicle itself.

[0060] Optionally, the step of performing spatiotemporal kinematic compensation processing on the initial relative pose reference sequence using the motion trajectory curvature features to generate the target relative pose result includes: Based on the curvature characteristics of the motion trajectory, determine whether the current motion state of the unmanned vehicle is in a high curvature turning state; When the judgment result indicates that the motion trajectory is in the large curvature turning state, a first weight adjustment coefficient is generated based on the curvature characteristics of the motion trajectory. The lateral offset in the initial relative pose reference sequence is amplified using the first weight adjustment coefficient to obtain transition pose data; Extract the chassis speed feedback quantity output in real time from the chassis speed measuring instrument of the unmanned vehicle to construct a state observation data structure; The transition pose data is introduced as an observation input into the state observation data structure for iterative convergence processing to perform spatiotemporal kinematic compensation processing, thereby generating the target relative pose result.

[0061] Preferably, before performing spatiotemporal kinematic compensation processing on the initial relative pose reference sequence using the motion trajectory curvature features, the motion trajectory curvature features, the initial relative pose reference sequence, the preset virtual three-dimensional coordinate system, the chassis speed feedback quantity output in real time by the chassis speedometer, and the current docking sampling timestamp corresponding to the unmanned vehicle performing docking control actions are first read. The motion trajectory curvature features are used to express the degree of trajectory curvature and curvature change trend of the unmanned vehicle within the historical sampling period; the initial relative pose reference sequence is used to express the visual pose estimation result of the unmanned vehicle relative to the docking target area; the preset virtual three-dimensional coordinate system is used to unify the spatial reference of the motion trajectory curvature features and the initial relative pose reference sequence; and the chassis speed feedback quantity is used to express the actual motion speed of the unmanned vehicle at the current docking sampling timestamp. The motion trajectory curvature features, the initial relative pose reference sequence, the preset virtual three-dimensional coordinate system, the chassis speed feedback, and the current docking sampling timestamp are aligned under the same time reference to form compensation input data for spatiotemporal kinematic compensation processing. The compensation input data continues to participate in subsequent large curvature turning state judgment, first weight adjustment coefficient generation, transition pose data generation, and state observation data structure construction, so that the large curvature turning state judgment, first weight adjustment coefficient generation, transition pose data generation, and state observation data structure construction can all be executed based on the same time reference and the same spatial reference.

[0062] Preferably, when determining whether the current motion state of the unmanned vehicle is in a high-curvature turning state based on the motion trajectory curvature features, the curvature value, curvature change trend, associated position state, and associated attitude state within the sampling time range adjacent to the current docking sampling timestamp are first read from the motion trajectory curvature features, and a pre-configured high-curvature turning judgment record is read; the high-curvature turning judgment record includes a preset curvature judgment threshold, a preset curvature duration condition, and a preset attitude change condition. The large curvature steering judgment record is pre-configured based on the minimum turning radius record of the unmanned vehicle, the allowable lateral offset range of the docking target area, and the steering response capability of the chassis actuators during the low-speed docking phase. The preset curvature judgment threshold is used to limit the range of curvature values ​​that reach the large curvature steering state judgment boundary. The preset curvature persistence condition is used to limit the number of adjacent sampling timestamps that need to be continuously covered when the curvature value reaches the preset curvature judgment threshold, in order to exclude inertial data fluctuations at a single sampling moment. The preset attitude change condition is used to limit the correspondence between the heading change direction and the curvature change trend in the associated attitude state, so that the large curvature steering state is not only determined by the curvature value alone, but also needs to correspond to the attitude change of the unmanned vehicle.

[0063] Preferably, when determining whether the current motion state of the unmanned vehicle is in the large curvature turning state, the curvature value within the sampling time range adjacent to the current docking sampling timestamp is compared with the preset curvature judgment threshold to obtain a curvature threshold comparison record. This record is then combined with the preset curvature persistence condition and the preset attitude change condition for joint determination. When the curvature value at multiple consecutive sampling timestamps reaches the preset curvature judgment threshold, and the curvature change trend is consistent with the heading change direction in the associated attitude state, the current motion state of the unmanned vehicle is determined to be in the large curvature turning state. When the curvature value reaches the preset curvature judgment threshold only at a single sampling timestamp, or when the curvature change trend does not correspond to the heading change direction in the associated attitude state, the current motion state of the unmanned vehicle is determined not to be in the large curvature turning state. Based on the above joint determination, the large curvature steering state can express the state of continuous bending motion of the unmanned vehicle during its approach to the docking target area, rather than the instantaneous disturbance of the inertial measurement unit; the determination result of the large curvature steering state continues to serve as the triggering basis for the generation of the first weight adjustment coefficient.

[0064] Preferably, when the judgment result indicates that the vehicle is in the large curvature turning state, the specific implementation process of generating the first weight adjustment coefficient based on the motion trajectory curvature features is as follows: First, the curvature value and curvature change trend corresponding to the current docking sampling timestamp are read from the motion trajectory curvature features, and the degree of deviation of the curvature value relative to the preset curvature judgment threshold is converted into a curvature deviation record; the curvature deviation record is used to express the degree to which the current trajectory curvature exceeds the judgment boundary corresponding to the preset curvature judgment threshold. Subsequently, the curvature deviation record is processed in correspondence with the chassis speed feedback amount to obtain a speed-related curvature record; the speed-related curvature record is used to express the difference in the impact of the same curvature value on the lateral offset correction requirement of the unmanned vehicle under different chassis speed feedback amounts. The velocity-related curvature record continues to undergo amplitude limiting and continuity processing according to the pre-configured weight adjustment configuration record to form the first weight adjustment coefficient. The weight adjustment configuration record includes a preset upper weight limit record, a preset lower weight limit record, and a weight change record for adjacent sampling timestamps. The preset upper weight limit record and the preset lower weight limit record are used to limit the value range of the first weight adjustment coefficient, and the weight change record for adjacent sampling timestamps is used to limit the change amplitude of the first weight adjustment coefficient under adjacent sampling timestamps. The first weight adjustment coefficient is a dimensionless adjustment record, used to subsequently proportionally adjust the lateral offset in the initial relative pose reference sequence, so that the adjustment amplitude of the lateral offset corresponds to the curvature of the current trajectory of the unmanned vehicle.

[0065] Preferably, when amplifying the lateral offset in the initial relative pose reference sequence using the first weight adjustment coefficient, the single-frame relative pose record corresponding to the current docking sampling timestamp is first read from the initial relative pose reference sequence, and the distance offset in the spatial depth direction, the lateral offset in the spatial lateral direction, the height offset in the spatial height direction, and the attitude deflection record are separated from the single-frame relative pose record. Subsequently, the first weight adjustment coefficient is applied to the lateral offset in the spatial lateral direction to obtain the compensated lateral offset; the compensated lateral offset is then recombined with the distance offset in the spatial depth direction, the height offset in the spatial height direction, and the attitude deflection record to form the transition pose data. The transition pose data does not change the distance offset in the spatial depth direction and the height offset in the spatial height direction obtained from the monocular visual image in the initial relative pose reference sequence. Instead, it compensates for the lateral offset in the spatial lateral direction, which is more susceptible to the lateral movement of the chassis under the large curvature steering state. This allows the transition pose data to reflect the lateral offset change of the unmanned vehicle relative to the docking target area during the bending approach process. The transition pose data continues to serve as the data source for the visual observation pose record in the state observation data structure.

[0066] Preferably, when extracting the real-time chassis speed feedback from the chassis velocimeter of the unmanned vehicle to construct the state observation data structure, the chassis speed sampling record output by the chassis velocimeter is first read according to the current docking sampling timestamp, and the sampling timestamp of the chassis speed sampling record is aligned so that the chassis speed sampling record and the transition pose data use the same sampling timestamp. The chassis speed sampling record after sampling timestamp alignment is used as the chassis speed feedback. The chassis speed feedback, together with the transition pose data, the motion trajectory curvature features, the adjacent single-frame relative pose records in the initial relative pose reference sequence, and the preset convergence judgment record, constitutes the state observation data structure. The preset convergence judgment record is pre-configured before the iterative convergence processing is executed based on the control refresh cycle of the chassis actuator, the image acquisition time of the monocular vision image, and the pose change range during the low-speed docking process of the unmanned vehicle, so that the preset convergence judgment record can provide the termination judgment basis for the iterative convergence processing. The state observation data structure is a time-varying data container that organizes multi-source state quantities for the same docking sampling timestamp. It includes visual observation pose records, chassis motion feedback records, curvature compensation records, predicted pose records, residual records, and convergence determination records. Specifically, the visual observation pose records originate from the transition pose data; the chassis motion feedback records originate from the chassis speed feedback; the curvature compensation records originate from the motion trajectory curvature features and the first weight adjustment coefficient; the predicted pose records are extended from the pose estimation results of the previous round in the iterative convergence process; the residual records express the deviation between the visual observation pose records and the predicted pose records; and the convergence determination records originate from the preset convergence determination records and are used to determine whether the iterative convergence process has ended.

[0067] Preferably, after the state observation data structure is constructed, the previous round of pose estimation results are first subjected to motion extension processing based on the chassis speed feedback and the sampling time interval between adjacent sampling timestamps to obtain the current round predicted pose record. In the first round of the iterative convergence processing, the previous round of pose estimation results are derived from the relative pose records of adjacent single frames located before the current docking sampling timestamp in the initial relative pose reference sequence. In non-first round processing, the previous round of pose estimation results are derived from the current round corrected pose record obtained in the previous iteration. The current round predicted pose record is used to express the relative pose at the current docking sampling timestamp, calculated solely based on the chassis motion feedback record without introducing the current visual observation pose record. Subsequently, the difference between the current round predicted pose record and the transition pose data is calculated to obtain the current round residual record. The current round residual record includes the distance residual in the spatial depth direction, the lateral residual in the spatial lateral direction, and the attitude deflection residual. The current round residual record continues to be weighted based on the motion trajectory curvature characteristics. When the motion trajectory curvature characteristics indicate that the current docking sampling timestamp is still in the high curvature turning state, the participation ratio of the lateral residual in the spatial lateral direction in the correction process is increased; when the motion trajectory curvature characteristics indicate that the current docking sampling timestamp tends to a low curvature motion state, the participation ratio of the lateral residual in the spatial lateral direction in the correction process is decreased. Through the above processing, the state observation data structure can organize the visual observation pose record, the chassis motion feedback record, and the curvature compensation record into the same round of iterative processing.

[0068] Preferably, when the transition pose data is introduced as an observation input into the state observation data structure for iterative convergence processing, the transition pose data is first written into the visual observation pose record in the state observation data structure, and the chassis speed feedback is written into the chassis motion feedback record in the state observation data structure. Subsequently, motion extension processing is performed on the previous round pose estimation result based on the chassis motion feedback record to obtain the current round predicted pose record. Then, pose correction processing is performed on the current round residual record between the visual observation pose record and the current round predicted pose record to obtain the current round corrected pose record. The current round corrected pose record is written back into the predicted pose record in the state observation data structure and participates in the motion extension processing as a new previous round pose estimation result in the next iteration. The above iterative convergence process ensures that the transition pose data is not directly output as the final result. Instead, the chassis speed feedback constrains the temporal continuity of the transition pose data, and the current wheel residual record constrains the visual observation deviation between the transition pose data and the current wheel predicted pose record, thereby reducing the lateral pose jump caused by the large curvature steering state round by round.

[0069] Preferably, the termination condition of the iterative convergence process is provided by the preset convergence determination record, which includes a residual change threshold record, an iteration count limit record, and a pose continuity determination record. The residual change threshold record is used to determine whether the residual record of the current iteration in two adjacent iterations has fallen below the allowable change range. The iteration count limit record is used to limit the maximum number of iterations that the iterative convergence process can execute. The pose continuity determination record is used to determine whether the current iteration corrected pose record and the relative pose record under the adjacent sampling timestamp conform to the continuous change trend during the low-speed docking process of the unmanned vehicle. When the current wheel residual record satisfies the residual change threshold record and the current wheel corrected pose record satisfies the pose continuity determination record, the iterative convergence process stops, and the current wheel corrected pose record is used as the target relative pose result. When the current wheel residual record does not satisfy the residual change threshold record, but the iteration round has not yet reached the upper limit of the iteration count, the next round of motion extension processing and pose correction processing continues. When the iteration round reaches the upper limit of the iteration count, the current wheel corrected pose record that satisfies the pose continuity determination record and has a lower current wheel residual record is selected as the target relative pose result. Through the joint constraints of the residual change threshold record, the upper limit of the iteration count record, and the pose continuity determination record, the target relative pose result can achieve a more suitable compensation state between visual observation bias and chassis motion continuity.

[0070] Preferably, the target relative pose result includes target distance offset, target lateral offset, and target attitude deflection; the target distance offset originates from the spatial depth direction correction result in the current wheel corrected pose record, the target lateral offset originates from the spatial lateral direction correction result in the current wheel corrected pose record, and the target attitude deflection originates from the attitude deflection correction result in the current wheel corrected pose record. The target relative pose result continues to retain the corresponding current docking sampling timestamp, the motion trajectory curvature feature, and the chassis speed feedback, so that when the target relative pose result is subsequently converted into a chassis traction control signal, it can simultaneously reflect the visual pose resolution result, the trajectory curvature state, and the actual chassis speed state. Therefore, the spatiotemporal kinematic compensation processing is not a simple smoothing of the initial relative pose reference sequence, but rather a layer-by-layer connection of the motion trajectory curvature feature, the first weight adjustment coefficient, the transition pose data, and the state observation data structure to form the target relative pose result that matches the current docking motion state.

[0071] Preferably, in a scenario, when the unmanned vehicle approaches the docking opening of the charging base station, if the motion trajectory curvature feature indicates that the unmanned vehicle has undergone continuous turning before approaching the docking opening, and the lateral offset in the initial relative pose reference sequence still mainly originates from the instantaneous observation of the monocular vision image, then the current motion state of the unmanned vehicle is determined to be in the large curvature turning state based on the motion trajectory curvature feature, and the first weight adjustment coefficient is generated based on the motion trajectory curvature feature. The first weight adjustment coefficient amplifies the lateral offset in the initial relative pose reference sequence to form the transition pose data. The transition pose data continues to be written into the state observation data structure together with the chassis speed feedback. The state observation data structure, through multiple rounds of motion extension processing, calculation of the current wheel residual record, and pose correction processing, ensures that the compensation result corresponding to the lateral offset in the spatial lateral direction remains temporally continuous with the chassis speed change, ultimately forming the target relative pose result. In this way, the target relative pose result can reduce the lateral lag of the single-frame visual pose in the large curvature steering state, so that the subsequent chassis traction control signal can be generated based on the relative pose that is closer to the current motion state.

[0072] Optionally, the step of converting it into a chassis traction control signal includes: The target relative pose result is analyzed to separate the relative distance deviation and relative heading angle deviation between the current position of the unmanned vehicle and the target docking position; A basic longitudinal traction force value is generated based on the relative distance deviation using a pre-configured proportional-integral-derivative controller. The relative heading angle deviation is multiplied by the pre-acquired chassis wheelbase parameters to generate a lateral moment offset value for correcting the heading. Based on the aforementioned longitudinal traction force value and the aforementioned lateral torque offset value, the chassis traction control signal is generated, which includes a left drive wheel torque command and a right drive wheel torque command.

[0073] Preferably, before analyzing the target relative pose result to separate the relative distance deviation and relative heading angle deviation between the current position of the unmanned vehicle and the target docking position, the target relative pose result, the preset virtual three-dimensional coordinate system, the target docking position record, and the unmanned vehicle current position record are read first. The target relative pose result is used to express the relative pose state after spatiotemporal kinematic compensation processing; the preset virtual three-dimensional coordinate system is used to define the spatial meaning of each directional component in the target relative pose result; the target docking position record is used to express the spatial position and docking orientation in the docking target area where the unmanned vehicle is allowed to complete docking control actions; and the unmanned vehicle current position record is used to express the spatial position of the unmanned vehicle relative to the preset virtual three-dimensional coordinate system at the current docking moment. The target docking position record includes a target depth position record and a target docking orientation record; the target depth position record is used to participate in the calculation of the difference in the relative distance deviation, and the target docking orientation record is used to participate in the calculation of the difference in the relative heading angle deviation. The target distance offset in the target relative pose result is calculated by differentiating it from the target depth position record to obtain the relative distance deviation. Similarly, the heading component in the target attitude deflection in the target relative pose result is calculated by differentiating it from the target docking orientation record to obtain the relative heading angle deviation. The relative distance deviation serves as the input for the proportional-integral-derivative (PID) controller to generate the base longitudinal traction force value, and the relative heading angle deviation serves as the input for generating the lateral moment offset value. This allows the target relative pose result to be decomposed into data required for longitudinal approach control and data required for heading correction control.

[0074] Preferably, before the relative distance deviation is input into the proportional-integral-derivative (PID) controller, it undergoes distance deviation sign determination and distance deviation amplitude normalization to form a distance deviation input record. The distance deviation sign determination distinguishes whether the current position of the unmanned vehicle relative to the target docking position is in a state of approaching or too close. The distance deviation amplitude normalization converts the relative distance deviations formed at different docking distances to the same traction force adjustment scale. The distance deviation input record retains the original value, sign direction, and corresponding sampling timestamp of the relative distance deviation. After the distance deviation input record is input into the PID controller, the PID controller does not directly output traction force according to a single proportional relationship. Instead, it first generates an instantaneous distance deviation component, a cumulative distance deviation component, and a distance change trend component based on the distance deviation input record. The instantaneous distance deviation component is used to subsequently generate a proportional traction component, the cumulative distance deviation component is used to subsequently generate an integral traction component, and the distance change trend component is used to subsequently generate a derivative traction component. This allows the basic longitudinal traction force value to simultaneously reflect the current distance deviation degree, the residual deviation during continuous approach, and the direction of change of the relative distance deviation.

[0075] Preferably, the proportional-integral-derivative controller includes, in terms of software execution structure, a distance deviation preprocessing module, a segmented proportional adjustment module, an integral memory adjustment module, a derivative suppression adjustment module, a low-speed traction limiting module, and a traction force output mapping module. The distance deviation preprocessing module receives the distance deviation input records and removes abnormal distance deviation input records according to the allowable distance range of the docking task, thereby forming a valid distance deviation input record that can participate in traction force adjustment. The segmented proportional adjustment module reads the valid distance deviation input records and generates the proportional traction component according to the distance interval where the instantaneous distance deviation component is located. The integral memory adjustment module accumulates residual deviations for multiple consecutive cumulative distance deviation components according to the sampling timestamp, thereby generating the integral traction component. The differential suppression adjustment module reads the distance change trend component under adjacent sampling timestamps and generates the differential traction component according to the distance change trend component. The low-speed traction limiting module reads the allowable traction range record for the low-speed docking stage and applies boundary constraints to the combination result of the proportional traction component, the integral traction component, and the differential traction component. The traction force output mapping module converts the combination result processed by the low-speed traction limiting module into the basic longitudinal traction force value. Through the aforementioned software execution structure, the proportional-integral-derivative controller can differ from the processing method that directly provides a fixed traction force based solely on the current distance deviation, thereby matching the basic longitudinal traction force value with the low-speed docking requirements of the unmanned vehicle when approaching the docking target area.

[0076] Preferably, when the proportional-integral-derivative (PID) controller generates the basic longitudinal traction force value based on the relative distance deviation, the distance deviation preprocessing module first reads the distance deviation input record and forms the valid distance deviation input record according to the allowable distance range record for the docking task. The allowable distance range record for the docking task is pre-configured based on the target docking position record and the low-speed traction capability of the chassis actuator, and continues to serve as the basis for the distance deviation preprocessing module to eliminate abnormal distance deviation input records. The valid distance deviation input record is then input to the segmented proportional adjustment module, which determines whether the instantaneous distance deviation component is in the far-distance approaching interval, the slow-speed approaching interval, or the end-approaching interval based on the segmented docking distance record, and generates the proportional traction component accordingly. The segmented docking distance record is pre-configured based on the target docking position record, the allowable distance range record for the docking task, and the low-speed traction capability of the chassis actuator, and continues to serve as the basis for the segmented proportional adjustment module to perform segmented processing. Subsequently, the integral memory adjustment module accumulates multiple consecutive cumulative distance deviation components according to the sampling timestamps to generate the integral traction component. The integral traction component is used to compensate for residual distance deviations caused by low-speed friction, changes in ground adhesion, or chassis response delays. The differential suppression adjustment module continues to read the distance change trend components under adjacent sampling timestamps to generate the differential traction component. The differential traction component is used to reduce the increasing trend of traction force caused by the proportional traction component and the integral traction component when the unmanned vehicle rapidly approaches the target docking position. The low-speed traction limiting module applies boundary constraints to the proportional traction component, the integral traction component, and the differential traction component. The traction force output mapping module then converts the boundary-constrained proportional traction component, the integral traction component, and the differential traction component into the basic longitudinal traction force value.

[0077] Preferably, the chassis wheelbase parameter is obtained in advance before participating in the generation of the lateral moment offset value. The chassis wheelbase parameter is derived from the unmanned vehicle chassis structure calibration record and the chassis drive wheel installation position record. The unmanned vehicle chassis structure calibration record is used to describe the geometric installation relationship of the drive wheels in the chassis actuator relative to the center of the unmanned vehicle chassis, and the chassis drive wheel installation position record is used to describe the equivalent arm relationship between the traction force application positions of the left and right drive wheels. For unmanned vehicles with left and right differential drive, the chassis wheelbase parameter can be determined by the lateral equivalent distance between the center positions of the left and right drive wheels; for unmanned vehicles with front and rear wheel sets, the chassis wheelbase parameter can be configured as the equivalent arm parameter for heading correction by combining the longitudinal distance between the front and rear wheel sets and the lateral equivalent distance between the left and right drive wheels. Therefore, the technical essence of the chassis wheelbase parameter is to convert the relative heading angle deviation into the geometric conversion basis required for the torque difference between the left and right drive wheels. The chassis wheelbase parameter is not a visual quantity temporarily inferred from the image, but rather pre-calibrated data corresponding to the structural dimensions of the chassis actuator and the installation position of the drive wheels.

[0078] Preferably, before multiplying the relative heading angle deviation with the pre-acquired chassis wheelbase parameter, the relative heading angle deviation is first subjected to heading deviation sign determination and heading deviation amplitude constraint processing to obtain a heading deviation input record. The heading deviation sign determination is used to distinguish whether the unmanned vehicle needs to be corrected to the left or right relative to the target docking position. The heading deviation amplitude constraint processing is used to limit the relative heading angle deviation exceeding the chassis's low-speed steering capability range to within the executable heading correction range. The heading deviation input record continues to retain the sign direction, amplitude, and corresponding sampling timestamp of the relative heading angle deviation. Subsequently, the relative heading angle deviation in the heading deviation input record is multiplied with the chassis wheelbase parameter to obtain the lateral moment offset value. The lateral moment offset value is used to express the offset amount that needs to be superimposed on the traction distribution of the left and right drive wheels to correct the relative heading angle deviation. By combining the relative heading angle deviation with the chassis wheelbase parameter, the lateral torque offset value can maintain a corresponding relationship with the chassis geometry of the unmanned vehicle, avoiding the problem of inconsistent dimensions caused by directly using the angle deviation as the difference in torque between the left and right drive wheels.

[0079] Preferably, after generating the lateral moment offset value, the basic longitudinal traction force value and the lateral moment offset value are first read, and the distribution direction of the lateral moment offset value between the left and right drive wheels is determined according to the sign direction in the heading deviation input record. When the heading deviation input record indicates that the unmanned vehicle needs to correct to the left, the lateral moment offset value is superimposed on the basic longitudinal traction force value in a manner that decreases on the left drive wheel and increases on the right drive wheel; when the heading deviation input record indicates that the unmanned vehicle needs to correct to the right, the lateral moment offset value is superimposed on the basic longitudinal traction force value in a manner that decreases on the right drive wheel and increases on the left drive wheel. The above distribution process yields the initial torque data for the left and right sides. The initial torque data for the left and right sides continues to participate in the drive wheel torque boundary constraint processing to avoid the lateral moment offset value causing excessive torque differences between the left and right drive wheels at the docking end.

[0080] Preferably, the drive wheel torque boundary constraint processing reads the left initial torque data, the right initial torque data, the chassis actuator's allowed torque range record, and the chassis speed feedback, and configures a corresponding drive wheel torque change constraint record based on the low-speed docking interval where the chassis speed feedback is located. The chassis actuator's allowed torque range record expresses the upper and lower limits of the torque allowed to be executed by the left and right drive wheels during docking control actions. The drive wheel torque change constraint record limits the torque change amplitude of the left and right drive wheels under adjacent sampling timestamps. After processing by the chassis actuator's allowed torque range record and the drive wheel torque change constraint record, the left initial torque data forms the left drive wheel torque command; after processing by the same chassis actuator's allowed torque range record and the same drive wheel torque change constraint record, the right initial torque data forms the right drive wheel torque command. Both the left and right drive wheel torque commands retain the sampling timestamp corresponding to the target relative pose result, enabling the chassis actuator to execute the docking control actions corresponding to the target relative pose result in chronological order.

[0081] Preferably, when generating the chassis traction control signal based on the basic longitudinal traction force value and the lateral moment offset value, the left drive wheel torque command, the right drive wheel torque command, the sampling timestamp corresponding to the target relative pose result, the relative distance deviation, the relative heading angle deviation, and the chassis wheelbase parameter are all written into the traction control command frame to form the chassis traction control signal. The traction control command frame is used to ensure that the chassis traction control signal not only includes the directly executable left drive wheel torque command and right drive wheel torque command, but also retains the relative pose deviation information and chassis geometric conversion information on which the left drive wheel torque command and right drive wheel torque command are generated; the relative pose deviation information includes the relative distance deviation and the relative heading angle deviation, and the chassis geometric conversion information includes the chassis wheelbase parameter. After the chassis traction control signal is sent to the chassis actuator, the chassis actuator adjusts the output torque of the left and right drive wheels according to the torque command of the left drive wheel and the torque command of the right drive wheel, respectively, so that the unmanned vehicle can maintain longitudinal proximity to the docking target area while completing the heading correction according to the lateral torque offset value.

[0082] Preferably, in one scenario, when the unmanned vehicle approaches the docking opening of the charging base station, the target relative pose result is first parsed into the relative distance deviation and the relative heading angle deviation. The relative distance deviation is input to the proportional-integral-derivative (PID) controller, and then processed sequentially by the distance deviation preprocessing module, the segmented proportional adjustment module, the integral memory adjustment module, the derivative suppression adjustment module, the low-speed traction limiting module, and the traction force output mapping module to form the basic longitudinal traction force value. The relative heading angle deviation is multiplied by the chassis wheelbase parameter to form the lateral torque offset value. Subsequently, the basic longitudinal traction force value is used to determine the basic longitudinal torque required for the left and right drive wheels to move forward together, and the lateral torque offset value is used to determine the torque difference between the left and right drive wheels. After the basic longitudinal traction force value and the lateral torque offset value are processed by drive wheel torque boundary constraints, the left drive wheel torque command and the right drive wheel torque command are formed. In this way, the chassis traction control signal can simultaneously respond to the distance deviation and heading deviation between the current position of the unmanned vehicle and the target docking position, so that the docking control action no longer depends on a single forward command or a single steering command, but is jointly determined by the target relative pose result, the proportional-integral-derivative controller, the chassis wheelbase parameter and the lateral moment offset value.

[0083] Optionally, before the step of generating the basic longitudinal traction force value based on the relative distance deviation using a pre-configured proportional-integral-derivative controller, the method further includes: Obtain the current ground adhesion coefficient of the environment in which the docking target area is located; Compare the current ground adhesion coefficient with a preset standard adhesion coefficient threshold; When the current ground adhesion coefficient is less than the standard adhesion coefficient threshold, the integral gain parameter in the pre-configured proportional-integral-derivative regulator is adjusted downwards. The derivative gain parameter in the pre-configured proportional-integral-derivative (PID) controller is adjusted upwards to obtain an updated PID controller. The relative distance deviation is input into the updated proportional-integral-derivative (PID) controller to trigger the step of generating a basic longitudinal traction force value based on the relative distance deviation according to the updated PID controller.

[0084] Preferably, before obtaining the current ground adhesion coefficient of the environment in which the docking target area is located, the following data are first read: the adhesion coefficient estimation data source record corresponding to the docking target area, the chassis speed feedback quantity, the drive wheel speed sampling record, the historical torque execution feedback record corresponding to the chassis traction control signal, and the sampling timestamp corresponding to when the unmanned vehicle performs the docking control action. The adhesion coefficient estimation data source record is used to limit the data sources participating in the current ground adhesion coefficient estimation. The chassis speed feedback quantity is used to express the actual moving speed of the chassis at the current docking time of the unmanned vehicle chassis center position. The drive wheel speed sampling record is used to express the rotation state of the left and right drive wheels at the same sampling timestamp. The historical torque execution feedback record is used to express the execution result of the chassis actuator on the left and right drive wheel torque commands at the previous sampling timestamp. The chassis speed feedback, the drive wheel speed sampling record, and the historical torque execution feedback record are aligned according to the same sampling timestamp to form a ground adhesion estimation input record. The ground adhesion estimation input record continues to participate in the estimation of the current ground adhesion coefficient, so that the current ground adhesion coefficient comes from the correspondence between the actual chassis movement speed, the drive wheel rotation state, and the execution result when the unmanned vehicle performs low-speed docking control actions in the docking target area.

[0085] Preferably, the specific implementation process for obtaining the current ground adhesion coefficient is as follows: First, based on the driving wheel rotation speed sampling record and the pre-calibrated driving wheel radius record, a theoretical rolling speed record of the driving wheels is generated; the theoretical rolling speed record of the driving wheels is used to express the actual chassis movement speed that the left and right driving wheels should correspond to when they are not in a driving wheel slipping state. Subsequently, the theoretical rolling speed record of the driving wheels and the chassis speed feedback quantity are processed to have the same dimension difference to form a driving wheel slipping observation record; the driving wheel slipping observation record is used to express the degree of deviation between the driving wheel rotation state and the actual chassis movement speed. The drive wheel slip observation record is further combined with the historical torque execution feedback record for traction response comparison processing to form a historical traction input strength record. The current ground adhesion coefficient is obtained based on the correspondence between the historical traction input strength record and the chassis speed feedback. The historical traction input strength record originates from the left drive wheel torque command, the right drive wheel torque command, and the historical torque execution feedback record at the previous sampling timestamp. The chassis speed feedback is used to express the actual chassis movement speed after the historical traction input strength record is applied. The current ground adhesion coefficient is used to express the traction conversion capability that the ground in the docking target area can provide to the drive wheels at the current docking moment. Through the above processing, the current ground adhesion coefficient can reflect the decrease in traction conversion capability caused by low-adhesion ground conditions, rather than being solely predetermined by the environmental type corresponding to the docking target area.

[0086] Preferably, the standard adhesion coefficient threshold is pre-configured before comparison. The standard adhesion coefficient threshold is derived from the ground type record of the docking target area, the traction demand record of the low-speed docking phase, and the traction response boundary record of the chassis actuator. The ground type record of the docking target area expresses the material type and surface condition of the ground in the docking target area. The traction demand record of the low-speed docking phase expresses the allowable low-speed traction range when the unmanned vehicle approaches the target docking position. The traction response boundary record of the chassis actuator expresses the allowable drive wheel traction output boundaries of the left and right drive wheels during docking control actions. After configuring the standard adhesion coefficient threshold based on the ground type record of the docking target area, the traction demand record of the low-speed docking phase, and the traction response boundary record of the chassis actuator, the standard adhesion coefficient threshold continues to serve as the comparison benchmark for whether the current ground adhesion coefficient meets the low-speed docking traction conditions. Both the standard adhesion coefficient threshold and the current ground adhesion coefficient are dimensionless adhesion capability characterization data, and they are compared using the same normalized scale to avoid directly comparing data with different dimensions.

[0087] Preferably, when comparing the current ground adhesion coefficient with a preset standard adhesion coefficient threshold, the current ground adhesion coefficient, the standard adhesion coefficient threshold, and the adhesion coefficient continuity determination record are first read. The adhesion coefficient continuity determination record is used to limit the number of sampling timestamps that need to be continuously covered when the current ground adhesion coefficient is lower than the standard adhesion coefficient threshold, so as to exclude instantaneous fluctuations in the chassis speed feedback or drive wheel speed sampling records at a single sampling moment. The current ground adhesion coefficient under multiple consecutive sampling timestamps is compared with the standard adhesion coefficient threshold respectively, and the comparison results are jointly determined with the adhesion coefficient continuity determination record; when the current ground adhesion coefficient under multiple consecutive sampling timestamps is less than the standard adhesion coefficient threshold, a low adhesion state determination record is formed; when the current ground adhesion coefficient is less than the standard adhesion coefficient threshold only at a single sampling timestamp, an adhesion instantaneous fluctuation determination record is formed. The low-attachment state determination record continues to be used to trigger the downward adjustment of the integral gain parameter and the upward adjustment of the derivative gain parameter. The attachment instantaneous fluctuation determination record continues to be used to maintain the parameter state of the proportional-integral-derivative controller at the current sampling timestamp, so that the proportional-integral-derivative controller will not be updated due to the attachment instantaneous fluctuation determination record.

[0088] Preferably, after the low adhesion state determination record is formed, the integral gain parameter in the pre-configured proportional-integral-derivative (PID) controller is adjusted downwards. The PID controller, in terms of software execution structure, includes a distance deviation preprocessing module, a segmented proportional adjustment module, an integral memory adjustment module, a derivative suppression adjustment module, a low-speed traction limiting module, and a traction output mapping module. The integral gain parameter belongs to the integral memory adjustment module and is used to control the participation intensity of multiple consecutive cumulative distance deviation components on the integral traction component. During the adjustment process, the low adhesion state determination record, the adhesion deviation degree of the current ground adhesion coefficient relative to the standard adhesion coefficient threshold, and the integral gain parameter currently used in the integral memory adjustment module are first read. Then, an integral gain adjustment amount is generated based on the adhesion deviation degree, and the integral gain parameter is adjusted downwards using this amount to obtain the adjusted integral gain parameter. The adjusted integral gain parameter is further written into the integral memory adjustment module, which reduces the integral traction component generated by the integral memory adjustment module in the low adhesion state, thereby reducing the trend of continuous increase in traction caused by the accumulation of residual distance deviation.

[0089] Preferably, when performing an upward adjustment process on the differential gain parameter in the pre-configured proportional-integral-derivative (PID) regulator, the low adhesion state determination record, the distance change trend component, and the differential gain parameter currently used in the differential suppression adjustment module are first read. The differential gain parameter belongs to the differential suppression adjustment module and is used to control the participation intensity of the distance change trend component on the differential traction component. The distance change trend component originates from the change in the relative distance deviation under adjacent sampling timestamps and is used to express the direction and magnitude of the distance deviation change when the unmanned vehicle approaches the target docking position. During the upward adjustment process, a differential gain upward adjustment amount is generated based on the degree of adhesion deviation of the current ground adhesion coefficient relative to the standard adhesion coefficient threshold, and the differential gain parameter is adjusted upward using this adjustment amount to obtain the adjusted differential gain parameter. The adjusted differential gain parameter is further written into the differential suppression adjustment module, which enhances the suppression effect on the increasing trend of the basic longitudinal traction force when the unmanned vehicle rapidly approaches the target docking position and the current ground adhesion coefficient is lower than the standard adhesion coefficient threshold. This makes the basic longitudinal traction force value more in line with the low-speed docking requirements under low adhesion conditions.

[0090] Preferably, after obtaining the adjusted integral gain parameter and the adjusted derivative gain parameter, the adjusted integral gain parameter is written into the integral memory adjustment module, and the adjusted derivative gain parameter is written into the derivative suppression adjustment module to obtain an updated proportional-integral-derivative (PID) controller. The updated PID controller still includes the distance deviation preprocessing module, the segmented proportional adjustment module, the integral memory adjustment module, the derivative suppression adjustment module, the low-speed traction limiting module, and the traction output mapping module; wherein the execution order of the distance deviation preprocessing module, the segmented proportional adjustment module, the low-speed traction limiting module, and the traction output mapping module remains unchanged. The integral memory adjustment module uses the adjusted integral gain parameter to generate the integral traction component, and the derivative suppression adjustment module uses the adjusted derivative gain parameter to generate the derivative traction component. Therefore, the updated PID controller does not reconstruct a new control flow, but rather replaces the parameters of the integral memory adjustment module and the derivative suppression adjustment module within the existing software execution structure based on the current ground adhesion coefficient.

[0091] Preferably, when the relative distance deviation is input into the updated proportional-integral-derivative (PID) controller, a distance deviation input record is first formed according to the relative distance deviation, and then the distance deviation input record is input into the distance deviation preprocessing module. The distance deviation preprocessing module reads the allowable distance range record for the docking task and performs validity screening on the distance deviation input record based on the allowable distance range record for the docking task to form a valid distance deviation input record. The allowable distance range record for the docking task originates from the allowable approach distance range corresponding to the target docking position and the executable distance boundary of the chassis actuator during the low-speed docking phase. The valid distance deviation input record is then input into the segmented proportional adjustment module, which generates a proportional traction component based on the instantaneous distance deviation component. Simultaneously, the cumulative distance deviation component in the valid distance deviation input record is input into the integral memory adjustment module, which generates the integral traction component based on the adjusted integral gain parameter. The distance change trend component in the valid distance deviation input record is input into the derivative suppression adjustment module, which generates the derivative traction component based on the adjusted derivative gain parameter. The proportional traction component, the integral traction component, and the differential traction component are further input into the low-speed traction limiting module for boundary constraints. The traction force output mapping module then converts the proportional traction component, the integral traction component, and the differential traction component after boundary constraints into the basic longitudinal traction force value.

[0092] Preferably, when the current ground adhesion coefficient is not less than the standard adhesion coefficient threshold, a normal adhesion state determination record is formed. This record is used to ensure the proportional-integral-derivative (PID) controller continues to use the pre-configured integral gain parameter and the pre-configured derivative gain parameter. The normal adhesion state determination record, the pre-configured integral gain parameter, and the pre-configured derivative gain parameter are continuously written into the PID controller parameter status record, which is used to mark whether the PID controller has undergone parameter updates at the current sampling timestamp. If the PID controller parameter status record indicates no parameter update, the relative distance deviation is input to the PID controller to trigger the generation of the basic longitudinal traction force value based on the relative distance deviation using the pre-configured PID controller. If the PID controller parameter status record indicates a parameter update, the relative distance deviation is input to the updated PID controller to trigger the generation of the basic longitudinal traction force value based on the relative distance deviation using the updated PID controller. In this way, the current ground adhesion coefficient can participate in the generation process of the basic longitudinal traction force value without changing the subsequent traction force output link.

[0093] Preferably, in one scenario, when the unmanned vehicle approaches the docking opening of the charging base station, if the ground adhesion estimation input record indicates an increased deviation between the theoretical rolling speed record of the drive wheels and the chassis speed feedback, and the historical traction input strength record formed after the historical torque execution feedback record undergoes traction response comparison processing indicates that the drive wheels have received traction input, then a current ground adhesion coefficient lower than the standard adhesion coefficient threshold is obtained based on the drive wheel slip observation record, the historical traction input strength record, and the chassis speed feedback. Subsequently, the current ground adhesion coefficient is compared with the standard adhesion coefficient threshold, and after forming the low adhesion state determination record, the integral gain parameter in the integral memory adjustment module is lowered, and the derivative gain parameter in the derivative suppression adjustment module is raised, resulting in the updated proportional-integral-derivative (PID) regulator. After the relative distance deviation is input into the updated proportional-integral-derivative (PID) controller, the integral memory adjustment module reduces the impact of the cumulative residual distance deviation on the integral traction component based on the adjusted integral gain parameter. The derivative suppression adjustment module enhances the suppression effect on the derivative traction component when the distance changes rapidly based on the adjusted derivative gain parameter. This makes the basic longitudinal traction force value more suitable for driving the chassis actuator to perform low-speed, continuous docking control actions under low adhesion conditions.

[0094] Optionally, after the step of causing the chassis actuator to perform a matching docking control action according to the chassis traction control signal, the method further includes: During the process of the chassis actuator performing the docking control action, the actual docking trajectory of the unmanned vehicle is extracted in real time; Obtain the pre-constructed theoretical docking trajectory; Calculate the dynamic following error value between the actual docking trajectory and the theoretical docking trajectory; Determine whether the dynamic following error value is greater than a preset safety tolerance threshold; When the dynamic following error value is greater than the safety tolerance threshold, an emergency braking command is generated; The emergency braking command is sent to the chassis actuator; The chassis actuator is controlled to stop performing the docking control action and enter the parking lock state.

[0095] Preferably, after the chassis actuator executes the docking control action according to the chassis traction control signal, the chassis traction control signal, the torque execution feedback record corresponding to the chassis actuator, the chassis speed feedback, the target relative pose result, the target docking position record, the preset virtual three-dimensional coordinate system, and the sampling timestamp corresponding to the unmanned vehicle executing the docking control action are read first. The chassis traction control signal is used to express the torque command of the left drive wheel and the torque command of the right drive wheel. The torque execution feedback record is used to express the execution status of the chassis actuator for the torque commands of the left drive wheel and the right drive wheel. The chassis speed feedback is used to express the actual moving speed of the chassis center position at the corresponding sampling timestamp. The target relative pose result is used to express the current spatial deviation state of the unmanned vehicle relative to the target docking position in the target docking position record. The target docking position record is used to provide the target spatial reference required for subsequent extraction of the actual docking trajectory and construction of the theoretical docking trajectory. The preset virtual three-dimensional coordinate system is used to unify the spatial reference of the target relative pose result, the actual docking trajectory, and the theoretical docking trajectory. The chassis traction control signal, the torque execution feedback record, the chassis speed feedback, the target relative pose result, and the target docking position record are aligned according to the same sampling timestamp to form an actual trajectory generation input record. The actual trajectory generation input record continues to participate in the extraction of the actual docking trajectory, so that the actual docking trajectory comes from the motion feedback after the chassis actuator performs the docking control action, rather than being inferred solely from the chassis traction control signal.

[0096] Preferably, when extracting the actual docking trajectory of the unmanned vehicle in real time, the displacement increment of the unmanned vehicle's chassis center position between adjacent sampling timestamps is calculated first based on the chassis speed feedback quantity in the input record generated by the actual trajectory and the time interval between adjacent sampling timestamps. Then, based on the heading state component in the target relative pose result and the left and right drive wheel execution difference in the torque execution feedback record, the displacement increment is oriented to obtain the actual trajectory point at the current sampling timestamp. The actual trajectory point includes the planar position component, heading state component, and sampling timestamp component in the preset virtual three-dimensional coordinate system. Multiple consecutive actual trajectory points are arranged in the order of sampling timestamps to form the actual docking trajectory. The actual docking trajectory continues to serve as the data basis for calculating the dynamic following error value, enabling subsequent error judgments to simultaneously reflect the spatial position deviation and heading state deviation of the unmanned vehicle.

[0097] Preferably, the theoretical docking trajectory is pre-constructed before the chassis actuator performs the docking control action. During construction, the target docking position record, the target relative pose result, the chassis wheelbase parameter, the low-speed docking phase traction requirement record, the docking target area boundary record, and the preset virtual three-dimensional coordinate system are read first. The target docking position record provides the target docking position that the unmanned vehicle should reach when completing the docking control action. The target relative pose result provides the relative distance deviation and relative heading angle deviation of the unmanned vehicle before it begins performing the docking control action. The chassis wheelbase parameter defines the geometric correspondence between the torque difference between the left and right drive wheels and the turning radius during heading correction. The low-speed docking phase traction requirement record defines the low-speed traction range of the unmanned vehicle as it approaches the target docking position. The docking target area boundary record defines the spatial range that the unmanned vehicle is allowed to pass through within the docking target area. The target docking position record, the target relative pose result, the chassis wheelbase parameter, the traction requirement record of the low-speed docking stage, and the docking target area boundary record are spatially normalized according to the preset virtual three-dimensional coordinate system to form a theoretical trajectory construction input record. The theoretical trajectory construction input record is then used to construct the theoretical docking trajectory.

[0098] Preferably, when constructing the theoretical docking trajectory based on the input record of the theoretical trajectory, the longitudinal approach segment is first determined according to the relative distance deviation in the target relative pose result, then the heading correction segment is determined according to the relative heading angle deviation in the target relative pose result and the chassis wheelbase parameter, and finally the end approach segment is determined according to the target docking position in the target docking position record and the boundary record of the docking target area. The longitudinal approach segment describes the trajectory part of the unmanned vehicle gradually approaching the target docking position from its current position, the heading correction segment describes the trajectory part of the unmanned vehicle eliminating the relative heading angle deviation during the approach process, and the end approach segment describes the trajectory part of the unmanned vehicle continuing to approach along the direction corresponding to the target docking position after entering the docking end. The longitudinal approach segment, the heading correction segment, and the end approach segment are spliced ​​together according to the sampling timestamp order, and position continuity processing and heading continuity processing are performed at the splicing position to obtain the theoretical docking trajectory; the theoretical docking trajectory includes multiple theoretical trajectory points, and each theoretical trajectory point includes a planar position component, a heading state component, and a sampling timestamp component in the preset virtual three-dimensional coordinate system.

[0099] Preferably, when acquiring the theoretical docking trajectory, the pre-constructed and stored theoretical docking trajectory and the theoretical trajectory construction input record are read, and multiple theoretical trajectory points in the theoretical docking trajectory are organized into a theoretical trajectory point sequence according to the sampling timestamp order. The theoretical trajectory point sequence and multiple actual trajectory points in the actual docking trajectory adopt the same preset virtual three-dimensional coordinate system, the same sampling timestamp reference, and the same trajectory point data structure. The trajectory point data structure includes a planar position component, a heading state component, and a sampling timestamp component, enabling the actual docking trajectory and the theoretical docking trajectory to be compared based on the trajectory point data structure with the same dimensions. In this way, the theoretical docking trajectory is not a simple straight-line reference trajectory, but a low-speed docking reference trajectory formed by combining the target relative pose result, the target docking position record, the chassis wheelbase parameter, the traction demand record of the low-speed docking stage, and the boundary record of the docking target area.

[0100] Preferably, when calculating the dynamic following error value between the actual docking trajectory and the theoretical docking trajectory, the actual docking trajectory, the theoretical docking trajectory, the actual trajectory points, the theoretical trajectory points, and the sampling timestamps are first read. Each actual trajectory point in the actual docking trajectory is time-matched with the theoretical trajectory point whose sampling timestamp is adjacent in the theoretical docking trajectory, forming a time-matched trajectory point pair. Then, the actual trajectory point and the theoretical trajectory point in the time-matched trajectory point pair are projected onto the same two-dimensional spatial plane, and the spatial position deviation between the actual trajectory point and the theoretical trajectory point is calculated. Combined with the heading deviation between the actual trajectory point and the theoretical trajectory point, a trajectory following deviation record is formed. The trajectory following deviation record includes the spatial position deviation, the heading deviation, and the corresponding sampling timestamp. The trajectory following deviation record is further sorted according to the sampling timestamps, and the dynamic following error value, which characterizes the degree of deviation in the current docking process, is extracted from the trajectory following deviation record. This allows the dynamic following error value to reflect the comprehensive following deviation of the unmanned vehicle in terms of spatial position and heading status.

[0101] Preferably, the safety tolerance threshold is preset before judgment. During setting, the following data are read: the boundary record of the docking target area, the target docking position record, the braking response record corresponding to the chassis actuator, the pose estimation error record corresponding to the target relative pose result, and the traction requirement record for the low-speed docking phase. The boundary record of the docking target area expresses the allowable space range occupied by the unmanned vehicle within the docking target area; the target docking position record expresses the target docking position when the unmanned vehicle completes the docking control action; the braking response record expresses the braking distance range formed by the chassis actuator from receiving the emergency braking command to stopping the docking control action; the pose estimation error record expresses the spatial error range formed by the target relative pose result during visual estimation; and the traction requirement record for the low-speed docking phase expresses the allowable low-speed traction range for the unmanned vehicle at the docking end. After configuring the safety tolerance threshold based on the boundary record of the docking target area, the target docking position record, the braking response record, the pose estimation error record, and the traction requirement record for the low-speed docking phase, the safety tolerance threshold continues to serve as a comparison benchmark for judging whether the dynamic following error value is within the allowable deviation range.

[0102] Preferably, the safety tolerance threshold is configured using the same spatial scale as the dynamic following error value, and the safety tolerance threshold does not directly use chassis torque, chassis speed feedback, or heading deviation as comparison objects. When configuring the safety tolerance threshold, the braking distance range corresponding to the braking response record, the spatial error range corresponding to the pose estimation error record, and the remaining space corresponding to the docking target area boundary record are converted to the same dimension to obtain the safety deviation allowable record; then, the safety tolerance threshold is generated based on the safety deviation allowable record and the low-speed docking stage traction demand record. The safety tolerance threshold is then compared with the dynamic following error value, ensuring that the comparison process is based on the same spatial scale, avoiding inconsistencies in technical logic caused by directly mixing and comparing the heading deviation, chassis speed feedback, and spatial distance deviation.

[0103] Preferably, when determining whether the dynamic following error value is greater than the safety tolerance threshold, the dynamic following error value, the safety tolerance threshold, and the trajectory error persistence determination record are first read. The trajectory error persistence determination record is used to limit the number of sampling timestamps that need to be continuously covered when the dynamic following error value is greater than the safety tolerance threshold, so as to exclude the instantaneous deviation of the actual trajectory point or the theoretical trajectory point under a single sampling timestamp. The dynamic following error value under multiple consecutive sampling timestamps is compared with the safety tolerance threshold, and the comparison result is jointly determined with the trajectory error persistence determination record; when the dynamic following error value under multiple consecutive sampling timestamps is greater than the safety tolerance threshold, a trajectory boundary violation determination record is formed; when the dynamic following error value is greater than the safety tolerance threshold only under a single sampling timestamp, a trajectory instantaneous deviation determination record is formed. The trajectory boundary violation determination record is still used to trigger the generation of the emergency braking command, and the trajectory instantaneous deviation determination record is still used to maintain the docking control action currently being performed by the chassis actuator.

[0104] Preferably, the emergency braking command is generated after the trajectory boundary violation determination record is formed. When generating the emergency braking command, the trajectory boundary violation determination record, the chassis traction control signal, the chassis speed feedback, the left drive wheel torque command, the right drive wheel torque command, and the current sampling timestamp are read first. Then, based on the trajectory boundary violation determination record, the control content that continues to increase traction in the chassis traction control signal is canceled, and the braking intensity control content is configured based on the chassis speed feedback, thus forming the emergency braking command. The emergency braking command includes a command sampling timestamp, traction cancellation content, braking intensity control content, and a parking lock determination record. The traction cancellation content is used to prevent the left drive wheel torque command and the right drive wheel torque command from continuing to increase in the original traction direction. The braking intensity control content is used to limit the braking intensity used when the chassis actuator performs the braking control action. The parking lock determination record is used to limit the determination conditions that the chassis actuator must meet when entering the parking lock state from the braking control action. The emergency braking command is then sent to the chassis actuator, enabling the chassis actuator to switch from performing the docking control action to performing the braking control action.

[0105] Preferably, when the emergency braking command is issued to the chassis actuator, the control command queue corresponding to the chassis actuator is first read. This control command queue is a pre-established command buffer data queue created when the chassis actuator receives the chassis traction control signal and the emergency braking command, used to organize the control commands to be executed according to the sampling timestamp order. The emergency braking command is written into the control command queue, and the sampling timestamp of the emergency braking command is compared with the sampling timestamp of the chassis traction control signal. When the sampling timestamp of the emergency braking command is later than the sampling timestamp of the chassis traction control signal, the emergency braking command is placed in the execution pending position in the control command queue, and the writing of new chassis traction control signals to the chassis actuator is stopped. When there are still unexecuted left drive wheel torque commands and right drive wheel torque commands in the control command queue, the unexecuted left drive wheel torque commands and right drive wheel torque commands are marked as invalid according to the emergency braking command, forming a torque command invalidation record. The torque command failure flag record continues to be written into the control command queue, causing the chassis actuator to perform braking control actions according to the emergency braking command, instead of continuing to perform docking control actions according to the chassis traction control signal that has not yet been executed.

[0106] Preferably, when controlling the chassis actuator to stop executing the docking control action and enter the parking lock state, the emergency braking command, the chassis speed feedback, the torque execution feedback record, and the parking lock determination record are first read. The parking lock determination record is used to limit the chassis speed conditions and torque execution conditions that the chassis actuator must meet when entering the parking lock state; the chassis speed conditions are derived from the chassis speed feedback, and the torque execution conditions are derived from the torque execution feedback record. If the chassis speed feedback satisfies the chassis speed conditions, and the torque execution feedback record indicates that the chassis actuator has stopped executing the left drive wheel torque command and the right drive wheel torque command, then a parking lock execution record is generated, and the chassis actuator is controlled to enter the parking lock state according to the parking lock execution record. The parking lock execution record is used to indicate that the chassis actuator has completed the braking control action and entered the parking lock state. The parking lock state is used to keep the unmanned vehicle stationary in the trajectory deviation scenario corresponding to the trajectory boundary judgment record, thereby preventing the chassis actuator from continuing to perform docking control actions, which would cause the dynamic following error value to increase further.

[0107] Preferably, in one scenario, when the unmanned vehicle approaches the docking opening of the charging base station, the chassis actuator executes the docking control action according to the chassis traction control signal, and forms the actual docking trajectory under multiple sampling timestamps. The pre-constructed theoretical docking trajectory is formed based on the target docking position record, the target relative pose result, the chassis wheelbase parameter, the traction demand record of the low-speed docking stage, and the docking target area boundary record, and is used to express the low-speed docking reference trajectory that the unmanned vehicle should follow in the docking target area. If multiple actual trajectory points in the actual docking trajectory continuously deviate from multiple theoretical trajectory points in the theoretical docking trajectory, and the dynamic following error value exceeds the safety tolerance threshold, a trajectory boundary violation determination record is formed; an emergency braking command is generated based on the trajectory boundary violation determination record, and the emergency braking command is sent to the chassis actuator, causing the chassis actuator to stop executing the docking control action and enter the parking lock state.

[0108] Optionally, the step of calculating the dynamic following error value between the actual docking trajectory and the theoretical docking trajectory includes: Within a preset two-dimensional spatial plane, the actual docking trajectory is discretized into a sequence of actual trajectory points; The theoretical docking trajectory is discretized into a sequence of theoretical trajectory points; Iterate through each actual trajectory point in the sequence of actual trajectory points; Calculate the Euclidean distance between the actual trajectory point and the nearest theoretical trajectory point in the sequence of theoretical trajectory points to obtain the set of local distance deviations; The maximum Euclidean distance value extracted from the set of local distance deviations is the dynamic following error value.

[0109] Preferably, before calculating the dynamic following error value between the actual docking trajectory and the theoretical docking trajectory, the preset virtual three-dimensional coordinate system, the target docking position, the target docking position record, the docking target area boundary record, the actual docking trajectory, and the theoretical docking trajectory are read first. The target docking position record is used to store the spatial position of the target docking position in the preset virtual three-dimensional coordinate system, and the docking target area boundary record is used to store the allowable passage range of the docking target area in the preset virtual three-dimensional coordinate system. The preset two-dimensional spatial plane is preset according to the horizontal coordinate component corresponding to the direction of movement of the unmanned vehicle chassis center position in the preset virtual three-dimensional coordinate system. The preset two-dimensional spatial plane is used to carry the planar position component when the unmanned vehicle approaches the target docking position in the docking target area; the target docking position is further used to limit the endpoint reference in the preset two-dimensional spatial plane, and the docking target area boundary record is further used to limit the allowable passage range in the preset two-dimensional spatial plane. Through the above-mentioned pre-set method, there is a spatial correspondence between the preset two-dimensional spatial plane and the preset virtual three-dimensional coordinate system, so that the subsequent actual docking trajectory and the theoretical docking trajectory can be calculated in Euclidean distance under the same spatial reference, instead of mixing the spatial coordinates in the image plane coordinates, chassis motion coordinates and the boundary record of the docking target area.

[0110] Preferably, when the preset two-dimensional spatial plane is pre-set, the relative distance deviation and relative heading angle deviation in the target relative pose result are also read, and the longitudinal approach direction in the preset two-dimensional spatial plane is determined based on the relative distance deviation, and the lateral deviation direction in the preset two-dimensional spatial plane is determined based on the relative heading angle deviation. The longitudinal approach direction is used to express the main movement direction when the center position of the unmanned vehicle chassis approaches the target docking position, and the lateral deviation direction is used to express the direction when the center position of the unmanned vehicle chassis deviates laterally relative to the target docking position. The longitudinal approach direction and the lateral deviation direction jointly participate in the coordinate axis configuration of the preset two-dimensional spatial plane, so that the preset two-dimensional spatial plane can directly express the planar deviation relationship between the actual docking trajectory and the theoretical docking trajectory during the low-speed docking process; the planar deviation relationship continues to participate in the subsequent calculation of the Euclidean distance between the actual trajectory point sequence and the theoretical trajectory point sequence.

[0111] Preferably, when discretizing the actual docking trajectory into a sequence of actual trajectory points within the preset two-dimensional spatial plane, multiple actual trajectory points in the actual docking trajectory, the planar position component, heading state component, and sampling timestamp component corresponding to each actual trajectory point are first read. The planar position component corresponding to each actual trajectory point is projected onto the preset two-dimensional spatial plane, and the heading state component and sampling timestamp component corresponding to each actual trajectory point are retained to form a sequence of actual trajectory points with a time order. Each actual trajectory point in the sequence corresponds to the chassis center position of the unmanned vehicle at one sampling timestamp. The sequence of actual trajectory points continues to participate in the retrieval of the nearest theoretical trajectory point, so that the dynamic following error value comes from the continuous motion feedback after the chassis actuator performs the docking control action, rather than from the separate estimation of the chassis traction control signal.

[0112] Preferably, when discretizing the theoretical docking trajectory into a sequence of theoretical trajectory points, multiple theoretical trajectory points, the planar position component, the heading state component, and the sampling timestamp component corresponding to each theoretical trajectory point are first read. The planar position component corresponding to each theoretical trajectory point is projected onto the preset two-dimensional space plane, and the heading state component and the sampling timestamp component corresponding to each theoretical trajectory point are retained. Then, they are arranged sequentially according to the sampling timestamp component corresponding to each theoretical trajectory point to form the sequence of theoretical trajectory points. Each theoretical trajectory point in the sequence of theoretical trajectory points corresponds to the planar position that the unmanned vehicle should reach in the low-speed docking reference trajectory. The sequence of theoretical trajectory points continues to serve as the comparison benchmark for the actual trajectory point sequence, enabling the actual trajectory point sequence and the theoretical trajectory point sequence to perform Euclidean distance calculations of the same dimension within the same preset two-dimensional space plane.

[0113] Preferably, when traversing each actual trajectory point in the actual trajectory point sequence, the planar position component, heading state component, and sampling timestamp component corresponding to the current actual trajectory point are first read, and candidate theoretical trajectory points are selected from the theoretical trajectory point sequence based on the sampling timestamp component corresponding to the current actual trajectory point. The candidate theoretical trajectory points originate from theoretical trajectory points in the theoretical trajectory point sequence that are adjacent to the current actual trajectory point in terms of sampling timestamp, and these candidate theoretical trajectory points are further used to participate in the selection of the nearest theoretical trajectory point. By selecting candidate theoretical trajectory points based on the sampling timestamp component, it is possible to avoid directly comparing the current actual trajectory point with non-corresponding theoretical trajectory points across long time periods, making the dynamic following error value more closely match the real-time following state of the unmanned vehicle during the docking control action.

[0114] Preferably, when calculating the Euclidean distance between the current actual trajectory point and the nearest theoretical trajectory point in the theoretical trajectory point sequence, the planar position components corresponding to the current actual trajectory point and the planar position components corresponding to each candidate theoretical trajectory point are first read in the preset two-dimensional space plane; then, the candidate Euclidean distances between the current actual trajectory point and each candidate theoretical trajectory point are calculated respectively to obtain a candidate Euclidean distance record; then, the candidate Euclidean distance with the smallest value is selected from the candidate Euclidean distance record, and the candidate theoretical trajectory point corresponding to the candidate Euclidean distance with the smallest value is taken as the nearest theoretical trajectory point. The candidate Euclidean distance between the nearest theoretical trajectory point and the current actual trajectory point is taken as the Euclidean distance corresponding to the current actual trajectory point, and the Euclidean distance corresponding to the current actual trajectory point is further written into the local distance deviation set, so that the local distance deviation set can record the degree of local deviation of each actual trajectory point relative to the theoretical trajectory point sequence.

[0115] Preferably, when forming the local distance deviation set, each actual trajectory point in the actual trajectory point sequence, the nearest theoretical trajectory point corresponding to each actual trajectory point, the Euclidean distance corresponding to each actual trajectory point, and the sampling timestamp component corresponding to each actual trajectory point are bound together to obtain a local distance deviation record. Multiple local distance deviation records are arranged according to the sampling timestamp component to form the local distance deviation set. The local distance deviation set not only records the magnitude of the Euclidean distance between the actual trajectory point sequence and the theoretical trajectory point sequence, but also records the sampling timestamp position where the Euclidean distance occurs. This allows for the location of actual trajectory points with significant deviations in the actual docking trajectory when extracting the dynamic following error value, and enables the local distance deviation set to continue serving as the data source for the dynamic following error value.

[0116] Preferably, when extracting the maximum Euclidean distance value from the local distance deviation set as the dynamic following error value, the process first iterates through multiple local distance deviation records in the local distance deviation set and reads the Euclidean distance from each record. After sorting the multiple Euclidean distances by size, the largest Euclidean distance is selected as the maximum Euclidean distance value, and this value is written into the dynamic following error record. The dynamic following error record includes the maximum Euclidean distance value, the actual trajectory point corresponding to the maximum Euclidean distance value, the closest theoretical trajectory point corresponding to the maximum Euclidean distance value, and the sampling timestamp component corresponding to the maximum Euclidean distance value. The dynamic following error record is then used to form the dynamic following error value, enabling the dynamic following error value to express the maximum planar deviation of the actual docking trajectory from the theoretical docking trajectory.

[0117] Preferably, the dynamic following error value is further compared with the safety tolerance threshold, which uses the same planar distance scale as the dynamic following error value. The dynamic following error value originates from the Euclidean distance in the preset two-dimensional spatial plane, and the safety tolerance threshold also uses the allowable planar deviation distance in the preset two-dimensional spatial plane as a comparison benchmark, thereby avoiding direct comparison of the heading angle, chassis speed, or chassis torque with the dynamic following error value. The comparison result between the dynamic following error value and the safety tolerance threshold is further used to determine whether the chassis actuator needs to stop executing the docking control action, and triggers the generation of the emergency braking command when the dynamic following error value is greater than the safety tolerance threshold, so that the local distance deviation set, the dynamic following error value, and the emergency braking command form a continuous technical process.

[0118] Preferably, in one scenario, when the unmanned vehicle approaches the docking opening of the charging base station, the preset two-dimensional spatial plane is pre-set according to the horizontal coordinate components corresponding to the moving direction of the unmanned vehicle chassis center position in the preset virtual three-dimensional coordinate system, and the target docking position stored in the target docking position record is used as the endpoint reference. The actual docking trajectory is discretized to form the actual trajectory point sequence, and the theoretical docking trajectory is discretized to form the theoretical trajectory point sequence. When a certain actual trajectory point in the actual trajectory point sequence deviates laterally from the theoretical trajectory point in the theoretical trajectory point sequence, the Euclidean distance between the actual trajectory point and the nearest theoretical trajectory point is written into the local distance deviation set. If the dynamic following error value formed by the maximum Euclidean distance value in the local distance deviation set is greater than the safety tolerance threshold, the dynamic following error value continues to participate in the generation of the emergency braking command, causing the chassis actuator to stop executing the docking control action and enter the parking lock state.

[0119] like Figure 3 As shown, this application provides an embodiment of a high-precision relative pose estimation and docking device for unmanned vehicles based on monocular vision. The device includes: The image acquisition module is used to acquire monocular vision images containing the docking target area collected by the monocular vision perception device installed at the front of the unmanned vehicle; The edge feature extraction module is used to perform multi-level edge feature extraction processing on the monocular vision image to obtain a multi-scale edge pixel cluster of the docking target region. The texture feature extraction module is used to perform multi-level texture feature extraction processing on the monocular vision image to obtain local texture gradient features. The three-dimensional spatial reconstruction module is used to perform three-dimensional spatial reconstruction processing on the multi-scale edge pixel cluster to obtain an initial three-dimensional point cloud structure. The pose resolution module is used to perform pose resolution processing on the initial three-dimensional point cloud structure using the local texture gradient features to obtain an initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system. The curvature feature extraction module is used to extract motion trajectory curvature features from the pose state data of the unmanned vehicle in continuous frames within the historical sampling period; The pose compensation and control module is used to perform spatiotemporal kinematic compensation processing on the initial relative pose reference sequence using the curvature features of the motion trajectory, generate the target relative pose result and convert it into a chassis traction control signal, so that the chassis actuator can perform matching docking control actions according to the chassis traction control signal.

[0120] like Figure 4 As shown, this is an embodiment of an electronic device according to this application. The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision as described in any one of this application.

[0121] like Figure 5 As shown, this is a computer-readable storage medium according to an embodiment of the present application. The computer-readable storage medium stores computer instructions, which are used to cause a computer to execute the high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision as described in any one of the present applications.

[0122] Figures 2-5 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A method for high-precision relative pose estimation of unmanned vehicle docking based on monocular vision, characterized in that, The high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision includes the following steps: Acquire monocular visual images containing the docking target area collected by a monocular visual perception device installed at the front of the unmanned vehicle; Perform multi-level edge feature extraction processing on the monocular vision image to obtain a multi-scale edge pixel cluster of the docking target region; Multi-level texture feature extraction processing is performed on the monocular vision image to obtain local texture gradient features; The multi-scale edge pixel cluster is subjected to three-dimensional spatial reconstruction processing to obtain an initial three-dimensional point cloud structure; The initial three-dimensional point cloud structure is subjected to pose analysis processing using the local texture gradient features to obtain an initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system. Extract motion trajectory curvature features from continuous frame pose state data of autonomous vehicles within a historical sampling period; The initial relative pose reference sequence is processed by spatiotemporal kinematic compensation using the curvature features of the motion trajectory to generate the target relative pose result and convert it into a chassis traction control signal, so that the chassis actuator can perform a matching docking control action according to the chassis traction control signal.

2. The monocular vision-based unmanned vehicle high-precision relative pose estimation docking method of claim 1, wherein, The monocular vision image is subjected to multi-level edge feature extraction processing to obtain a multi-scale edge pixel cluster of the docking target region; The steps of performing multi-level texture feature extraction processing on the monocular vision image to obtain local texture gradient features include: The monocular visual image is mapped from the reference three primary color color space to the hue, saturation and brightness color space to obtain the converted image data; A pre-configured spatial smoothing mask is invoked to perform noise filtering on the transformed image data, resulting in smoothed image data; Extract the pixel brightness abrupt change regions from the smoothed image data; The pixel brightness abrupt change regions are aggregated to perform multi-level edge feature extraction processing, thereby obtaining a multi-scale edge pixel cluster of the docking target region; Extract the grayscale change rate between adjacent pixels in the smoothed image data to generate a basic texture gradient field; The smoothed image data is divided into multiple local spatial blocks of different scales; Within the multiple local spatial blocks of different scales, the direction and amplitude of the basic texture gradient field are statistically aggregated to perform multi-level texture feature extraction processing, thereby obtaining the local texture gradient features. 3.The monocular vision based high-precision relative pose estimation docking method of an unmanned vehicle according to claim 1, wherein, The steps of performing three-dimensional spatial reconstruction processing on the multi-scale edge pixel cluster to obtain an initial three-dimensional point cloud structure, and performing pose analysis processing on the initial three-dimensional point cloud structure using the local texture gradient features to obtain an initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system include: Obtain the pre-calibrated set of internal calibration parameters of the monocular vision sensing device; The internal calibration parameter set is used to perform inverse projection calculation on the two-dimensional pixel coordinates in the multi-scale edge pixel cluster to generate a first three-dimensional spatial ray set; Calculate the spatial intersection points of the first three-dimensional spatial ray set within the preset depth constraint range; Spatial topology aggregation is performed on the spatial intersections to perform the three-dimensional spatial reconstruction process, resulting in the initial three-dimensional point cloud structure. The estimated normal vector of the docking target surface is determined based on the local texture gradient features. The initial 3D point cloud structure is projected onto the projection plane corresponding to the normal vector estimate to obtain the relative translation vector and the relative rotation feature set. The pose analysis process is performed based on the relative translation vector and the relative rotation feature set to obtain the initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system. 4.The monocular vision based high-precision relative pose estimation docking method of an unmanned vehicle according to claim 1, wherein, The step of extracting motion trajectory curvature features from the continuous frame pose state data of the unmanned vehicle within the historical sampling period includes: The inertial measurement unit inside the unmanned vehicle records the acceleration time series and angular velocity time series during the historical sampling period. The continuous frame pose state data includes the acceleration time series and angular velocity time series. Perform a first time accumulation operation on the acceleration time series to obtain the velocity time series; Perform a second time accumulation operation on the velocity time series to obtain the displacement time series; Based on the displacement time series and the angular velocity time series, the continuous frame pose state data is generated; Determine the pose points of three adjacent frames in the continuous frame pose state data; Calculate the spatial dimensions of the circumcircle of the spatial triangle formed by the pose points of the three adjacent frames; The curvature characteristics of the motion trajectory are calculated based on the circumscribed circle space dimensions. 5.The monocular vision based high-precision relative pose estimation docking method of an unmanned vehicle according to claim 1, wherein, The step of performing spatiotemporal kinematic compensation processing on the initial relative pose reference sequence using the curvature features of the motion trajectory to generate the target relative pose result includes: Based on the curvature characteristics of the motion trajectory, determine whether the current motion state of the unmanned vehicle is in a high curvature turning state; When the judgment result indicates that the motion trajectory is in the large curvature turning state, a first weight adjustment coefficient is generated based on the curvature characteristics of the motion trajectory. The lateral offset in the initial relative pose reference sequence is amplified using the first weight adjustment coefficient to obtain transition pose data; Extract the chassis speed feedback quantity output in real time from the chassis speed measuring instrument of the unmanned vehicle to construct a state observation data structure; The transition pose data is introduced as an observation input into the state observation data structure for iterative convergence processing to perform spatiotemporal kinematic compensation processing, thereby generating the target relative pose result.

6. The high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision as described in claim 1, characterized in that, The step of converting it into a chassis traction control signal includes: The target relative pose result is analyzed to separate the relative distance deviation and relative heading angle deviation between the current position of the unmanned vehicle and the target docking position; A basic longitudinal traction force value is generated based on the relative distance deviation using a pre-configured proportional-integral-derivative controller. The relative heading angle deviation is multiplied by the pre-acquired chassis wheelbase parameters to generate a lateral moment offset value for correcting the heading. Based on the aforementioned longitudinal traction force value and the aforementioned lateral torque offset value, the chassis traction control signal is generated, which includes a left drive wheel torque command and a right drive wheel torque command.

7. The high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision as described in claim 6, characterized in that, Before the step of generating the basic longitudinal traction force value based on the relative distance deviation using a pre-configured proportional-integral-derivative controller, the method further includes: Obtain the current ground adhesion coefficient of the environment in which the docking target area is located; Compare the current ground adhesion coefficient with a preset standard adhesion coefficient threshold; When the current ground adhesion coefficient is less than the standard adhesion coefficient threshold, the integral gain parameter in the pre-configured proportional-integral-derivative regulator is adjusted downwards. The derivative gain parameter in the pre-configured proportional-integral-derivative (PID) controller is adjusted upwards to obtain an updated PID controller. The step of inputting the relative distance deviation into the updated proportional-integral-derivative (PID) controller to trigger the generation of a basic longitudinal traction force value based on the relative distance deviation according to the updated PID controller.

8. A high-precision relative pose estimation and docking device for unmanned vehicles based on monocular vision, characterized in that, The high-precision relative pose estimation and docking device for unmanned vehicles based on monocular vision includes: The image acquisition module is used to acquire monocular vision images containing the docking target area collected by the monocular vision perception device installed at the front of the unmanned vehicle; The edge feature extraction module is used to perform multi-level edge feature extraction processing on the monocular vision image to obtain a multi-scale edge pixel cluster of the docking target region. The texture feature extraction module is used to perform multi-level texture feature extraction processing on the monocular vision image to obtain local texture gradient features. The three-dimensional spatial reconstruction module is used to perform three-dimensional spatial reconstruction processing on the multi-scale edge pixel cluster to obtain an initial three-dimensional point cloud structure. The pose resolution module is used to perform pose resolution processing on the initial three-dimensional point cloud structure using the local texture gradient features to obtain an initial relative pose reference sequence located in a preset virtual three-dimensional coordinate system. The curvature feature extraction module is used to extract motion trajectory curvature features from the pose state data of the unmanned vehicle in continuous frames within the historical sampling period; The pose compensation and control module is used to perform spatiotemporal kinematic compensation processing on the initial relative pose reference sequence using the curvature features of the motion trajectory, generate the target relative pose result and convert it into a chassis traction control signal, so that the chassis actuator can perform matching docking control actions according to the chassis traction control signal.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the high-precision relative pose estimation and docking method for unmanned vehicles based on monocular vision as described in any one of claims 1 to 7.

Images