An AGV multi-sensor false obstacle elimination method and device based on vehicle body micro-disturbance fingerprint characteristics

Abstract

The application discloses an AGV multi-sensor false obstacle elimination method and device based on a vehicle body micro-disturbance fingerprint feature, relates to the general image data processing field, and has the advantages that the method comprises the following steps: based on the motion state parameters of an AGV vehicle body and a detected suspected obstacle region, the displacement target quantity of an observation baseline is calculated; the observation baseline is controlled to generate a micro-disturbance relative to the suspected obstacle region, and detection contrast data of the same suspected obstacle region is acquired; the detection contrast data is inversed to the same viewpoint coordinate system, and parallax displacement quantity and reflection intensity ratio are obtained; geometric residual and reflection residual are calculated, and the disturbance fingerprint feature of the suspected obstacle region is determined; the template similarity between the disturbance fingerprint feature and a preset entity reference template is calculated; when the template similarity is greater than a preset limit, it is determined that the suspected obstacle region exists a real physical entity, and corresponding obstacle avoidance control instructions are generated. By implementing the application, the missing detection and wrong detection probability of obstacles can be reduced.

Description

Technical Field

[0001] This application relates to the general field of image data processing, and in particular to a method and apparatus for removing false obstacles in AGVs using multi-sensor technology based on the fingerprint features of micro-disturbances on the vehicle body. Background Technology

[0002] In modern warehousing and logistics systems, automated guided vehicles (AGVs) are the core carriers of material flow. During the autonomous navigation operation of AGVs, they need to continuously detect the surrounding environment to avoid obstacles on the travel path, ensure the safety of equipment and goods, and maintain the efficient operation of workshop logistics.

[0003] In related technologies, the environmental perception scheme adopts the method of synchronously collecting multimodal environmental data to establish a static environmental grid. The control module extracts the geometric contour and reflection intensity of the target within the grid, and calculates the static confidence score by combining it with a preset weight matrix. When the static confidence score of a specific area exceeds the set threshold in multiple consecutive observation periods, the system confirms the existence of physical obstacles and issues a path replanning instruction.

[0004] However, due to its reliance on passive acquisition and static cross-validation, the system struggles to differentiate the real physical properties from the isomorphic static features when faced with visual interference sources that produce similar static feedback and real low-lying objects, resulting in a high probability of missed or false detections. Summary of the Invention

[0005] This application provides a method and apparatus for eliminating false obstacles in AGVs based on the fingerprint features of micro-disturbances on the vehicle body, which is used to reduce the probability of missed or false detection of obstacles.

[0006] Firstly, this application provides a method for eliminating false obstacles in an AGV using multiple sensors based on the micro-disturbance fingerprint features of the vehicle body. This method is applied to an obstacle recognition device. The method includes: calculating the target displacement of the observation baseline between the environmental detection component and the suspected obstacle area based on the motion state parameters of the AGV body and the detected suspected obstacle area; controlling the observation baseline to generate a micro-disturbance relative to the suspected obstacle area, and acquiring detection comparison data of the same suspected obstacle area before and after the micro-disturbance; combining the target displacement of the observation baseline, inverting the detection comparison data to the same viewpoint coordinate system to obtain the disparity displacement and reflection intensity ratio of the suspected obstacle area; calculating the geometric residual and reflection residual based on the disparity displacement and reflection intensity ratio to determine the disturbance fingerprint features of the suspected obstacle area; calculating the template similarity between the disturbance fingerprint features and a preset entity reference template; and determining that a real physical entity exists in the suspected obstacle area when the template similarity is greater than a preset limit, and generating corresponding obstacle avoidance control commands.

[0007] In the above embodiments, the obstacle recognition device constructs a perturbation fingerprint by acquiring the multidimensional dynamic change features of the target generated by environmental stimuli, which can distinguish between static entities and false targets in the environment, thereby improving the accuracy of the recognition strategy and obstacle avoidance planning.

[0008] In some embodiments of the first aspect, the step of calculating geometric residuals and reflection residuals by combining disparity displacement and reflection intensity ratio to determine perturbation fingerprint features of suspected obstacle regions specifically includes: extracting depth constraints defined by the displacement target amount of the observation baseline, determining geometric residuals characterizing the consistency of three-dimensional structures by combining disparity displacement; calculating a mapping function of reflection intensity ratio as a function of viewing angle change with the viewing angle change before and after micro-perturbation, determining reflection residuals based on the deviation of reflection intensity ratio and mapping function; and constructing perturbation fingerprint features based on geometric residuals and reflection residuals.

[0009] In the above embodiments, the obstacle recognition device can extract the deep physical properties of the target and form a highly quantified recognition vector by calculating the multi-view residuals in the geometric and reflection dimensions, thereby enhancing the recognition robustness of the device in multi-material scenarios.

[0010] In conjunction with some embodiments of the first aspect, in some embodiments, after the steps of extracting the depth constraint defined by the displacement target amount of the observation baseline and determining the geometric residual characterizing the consistency of the three-dimensional structure in combination with the disparity displacement amount, the method further includes: dividing the suspected obstacle region into a first component consistent with the depth constraint and a second component inconsistent with the depth constraint according to the sign polarity of the disparity displacement amount; when the proportion of the second component is greater than a preset proportion threshold, determining the suspected obstacle region as a virtual image of the reflective surface and removing it; when the proportion of the second component is not greater than the preset proportion threshold, correcting the reflection intensity ratio with the reflection intensity of the first component.

[0011] In the above embodiments, the obstacle recognition device can strip away interfering features in suspected targets and remove them when they are determined to be virtual images of reflective surfaces. At the same time, it uses the reflection intensity of the real components to correct the extraction results, thereby improving the overall recognition accuracy in environments with mirror refraction.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, the step of generating a micro-disturbance of the displacement target amount relative to the suspected obstacle area by the control observation baseline specifically includes: calculating the steering wheel deflection amplitude and the duration of the disturbance based on the displacement target amount and the driving speed in the motion state parameters; and generating a steering wheel deflection command by combining the steering wheel deflection amplitude and the duration of the disturbance to generate a micro-disturbance in the AGV body.

[0013] In the above embodiments, the obstacle recognition device can dynamically calculate control commands based on real-time operating conditions to achieve smooth trajectory deflection, generating the required observation parallax without affecting normal cruise and anti-overturning.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, the step of controlling the observation baseline to generate a micro-perturbation of the displacement target amount relative to the suspected obstacle area specifically includes: obtaining the load center of gravity height and wheelbase of the AGV vehicle body; determining the upper limit of lateral acceleration of the AGV vehicle body based on the load center of gravity height and wheelbase; determining the movement speed and movement stroke of the motion mechanism carrying the environmental detection component based on the upper limit of lateral acceleration and the displacement target amount; controlling the motion mechanism to move relative to the AGV vehicle body according to the movement speed and movement stroke, so as to generate a micro-perturbation of the environmental detection component.

[0015] In the above embodiments, the obstacle recognition device can control the independent motion mechanism to complete the relative movement of the probe, ensuring that the detection operation is completed without being restricted by the kinematic constraints of the vehicle chassis under ultra-high or ultra-heavy load conditions.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, before the step of calculating the template similarity between the perturbation fingerprint features and the preset entity reference template, the method further includes: determining an upper limit value of the parallax displacement based on the target displacement of the observation baseline and the nearest distance from the suspected obstacle area to the environmental detection component; when the parallax displacement is greater than the upper limit value, determining the suspected obstacle area as a dynamic target and generating a corresponding dynamic obstacle avoidance command.

[0017] In the above embodiments, the obstacle recognition device can filter non-static targets based on excessive displacement and execute response logic independently, avoiding the business risk of misjudging moving personnel or driving equipment as static interference debris.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, after determining the suspected obstacle area as a dynamic target and generating a corresponding dynamic obstacle avoidance command when the parallax displacement is greater than the upper limit value, the method further includes: determining the dynamic target's own displacement component based on the parallax displacement and the upper limit value; determining the dynamic target's approaching or moving away state relative to the AGV vehicle body based on the sign polarity of the own displacement component and the displacement direction of the observation baseline; generating a braking command to replace the dynamic obstacle avoidance command when the dynamic target is in the approaching state; and generating a deceleration observation command to replace the dynamic obstacle avoidance command when the dynamic target is in the moving away state.

[0019] In the above embodiments, the obstacle recognition device can combine dynamic component calculations to determine collision risks and issue different levels of deceleration actions accordingly, thus ensuring the safety of the working environment while maintaining the efficiency of the workshop's logistics operations.

[0020] In a second aspect, embodiments of this application provide an obstacle recognition device, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the obstacle recognition device to perform the method described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on an obstacle recognition device, cause the obstacle recognition device to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on an obstacle recognition device, cause the obstacle recognition device to perform the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the obstacle recognition device provided in the second aspect, the computer storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the methods provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0025] 1. By employing a technique that calculates the displacement target based on the AGV vehicle's motion state parameters and suspected obstacle areas, and controls the generation of micro-disturbances to obtain detection comparison data, it is possible to obtain the three-dimensional feature changes of the target as it moves with the observation coordinate system. This solves the technical problem in existing technologies where passive static filtering cannot distinguish between real obstacles and false obstacles caused by ambient light and shadow, and achieves accurate separation and stable removal of physical entities and optical reflection interference features.

[0026] 2. By employing techniques to extract depth constraints to determine geometric residuals and to determine reflection residuals using a parallax change mapping function, perturbation fingerprints with multi-modal representation capabilities can be constructed through multi-dimensional physical-level feature quantization. This solves the problem of single-feature judgment failure caused by the diversity of reflectivity and materials in industrial scenarios, and achieves high-confidence assessment of the authenticity of stationary obstacles under complex background interference.

[0027] 3. By employing a technique that calculates the steering wheel deflection amplitude and disturbance duration based on the driving speed to control micro-disturbances, a smooth and controllable observation baseline offset can be generated while ensuring vehicle dynamics stability and load safety. This solves the risks of trajectory deviation and material slippage that may be caused by fixed-amplitude disturbance operations, and achieves real-time dynamic adaptive detection without affecting normal cruise operations. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of a topology of the obstacle recognition device in the embodiments of this application;

[0029] Figure 2 This is a flowchart illustrating an AGV multi-sensor false obstacle removal method based on vehicle body micro-disturbance fingerprint features in an embodiment of this application.

[0030] Figure 3 This is another flowchart illustrating the AGV multi-sensor false obstacle removal method based on vehicle body micro-disturbance fingerprint features in this application embodiment;

[0031] Figure 4 This is a schematic diagram of the physical structure of an obstacle recognition device in the embodiments of this application. Detailed Implementation

[0032] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0033] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0034] In the specific implementation scenarios of industrial warehousing and workshop production, the workshop environment is full of complex optical and physical interference sources such as skylight light and shadow, strobe lights, floor reflections, and oil stains on the floor. When the equipment is in normal patrol or standby mode, the environmental detection system can easily misidentify these non-physical interferences as obstacles blocking passage.

[0035] To fundamentally distinguish between physically existing low-lying debris and reflective false obstacles attached to environmental surfaces, this application introduces a controllable micro-perturbation strategy. This strategy applies minute and transient spatial displacements to the vehicle or onboard sensor components through a low-level motion mechanism, thereby altering the observation baseline—the spatial vector connecting the sensor's optical center and the target point.

[0036] During this displacement process, the real physical entity is attached to the ground and maintains a fixed structure in space. As the observation baseline shifts or deflects, its surface will inevitably generate corresponding parallax displacement during multi-time-series acquisitions, exhibiting characteristics of continuous depth texture and stable physical reflection intensity. In contrast, the light spot or mirror virtual image depends on changes in the illumination angle, lacks actual physical depth structure, and its spatial distribution and reflection patterns under the influence of disturbances exhibit a nonlinear and disordered jump state.

[0037] By calculating the three-dimensional geometric residuals and surface reflection residuals of the key parameters before and after micro-disturbance, the degree of fluctuation is mathematically normalized and mapped to generate a disturbance fingerprint feature that reflects the multi-dimensional characteristics of the target's essence. Using this disturbance fingerprint feature for high-dimensional matching with pre-defined standard model data, the obstacle recognition device can accurately eliminate false environmental interference signals without consuming excessive floating-point computing resources, ensuring efficient and stable operation of automated guidance equipment in complex workshop environments.

[0038] For example, when faced with a large area of ​​water stains reflecting light on the ground surface, its outline and signal-to-noise ratio will fluctuate greatly after being stimulated by micro-motion. After fingerprint value comparison, it is confirmed that the target does not have real physical blocking attributes. The device ignores it and maintains the current driving speed parameters.

[0039] The system architecture connection method provided in this embodiment is described below. Please refer to... Figure 1 This is a schematic diagram of a topology architecture of the obstacle recognition device in the embodiments of this application.

[0040] like Figure 1 As shown, the system architecture mainly consists of three core components that work together: the suspected obstacle area 10, the AGV vehicle body 11, and the obstacle recognition device 12.

[0041] The suspected obstacle area 10 is a candidate detection range locked by the scan. The objective object attributes enclosed within this area are manifested as a real physical entity 101 with obstructive properties, or a virtual image 102 of a reflective surface induced by complex lighting and reflective surfaces in the workshop. The environmental detection component 112 continuously outputs obstacle area information to the obstacle recognition device 12 for the suspected obstacle area 10.

[0042] The AGV body 11 serves as the underlying physical carrier for the system's operation, and its interior is equipped with a motion mechanism / steering wheel 111 and an environmental detection component 112. The motion mechanism / steering wheel 111 receives drive signals from the underlying motor to execute the vehicle's trajectory movement or attitude deflection within the workspace. The environmental detection component 112 is composed of a series of multi-modal sensors, which are controlled to emit detection reference signals to the surrounding working environment and receive surface-reflected echoes. During operation, the AGV body 11 reports the real-time collected motion status parameters to the obstacle recognition device 12.

[0043] The obstacle recognition device 12 serves as the computational center for feature extraction and obstacle avoidance decision-making. Internally, it is logically divided into a displacement target quantity calculation module 121, a micro-disturbance control and acquisition module 122, a data coordinate inversion module 123, a disturbance fingerprint feature construction module 124, and a template matching and obstacle avoidance control module 125. Variable transfer between these computational modules is accomplished via an internal data bus.

[0044] The displacement target calculation module 121 receives motion state parameters from the AGV body 11 and obstacle area information from the environmental detection component 112 in parallel. Based on the above information, it calculates and outputs the displacement target of the observation baseline of the environmental detection component, and then transmits the displacement target to the next level module.

[0045] The micro-disturbance control and acquisition module 122 retrieves the displacement target amount and sends specific deflection / micro-motion commands to the motion mechanism / steering wheel 111 inside the AGV body 11 to control the observation baseline to generate micro-disturbances. Simultaneously, it receives multi-time series detection comparison data captured by the environmental detection component 112 during the application of the disturbance.

[0046] The data coordinate inversion module 123 receives the detection comparison data and uses the inverse projection matrix to invert the scattered data generated during the micro-disturbance to the same coordinate system, thereby extracting the parallax displacement and reflection intensity ratio of the suspected obstacle area.

[0047] The perturbation fingerprint feature construction module 124 calculates geometry and reflection residuals based on the parsed parallax displacement and reflection intensity ratio, and integrates multi-dimensional quantitative indicators to determine the perturbation fingerprint features of the target.

[0048] The template matching and obstacle avoidance control module 125 acquires the disturbance fingerprint features and identifies the authenticity of the suspected obstacle area by calculating the template similarity between feature vectors. When the target is confirmed to be a real physical entity, the module generates a corresponding obstacle avoidance control command and sends the command back to the control end of the AGV body 11 to execute the action intervention. This system architecture achieves a processing link combining detection and feature extraction through the closed-loop data interaction of the above components and modules, improving the system's identification accuracy in complex interference environments.

[0049] The method provided in this embodiment is described in detail below. Please refer to [link / reference]. Figure 2 This is a flowchart illustrating an AGV multi-sensor false obstacle removal method based on vehicle body micro-disturbance fingerprint features in an embodiment of this application.

[0050] S201. Based on the motion state parameters of the AGV vehicle body and the detected suspected obstacle area, calculate the target displacement of the observation baseline between the environmental detection component and the suspected obstacle area.

[0051] Among them, the AGV vehicle body refers to the transport equipment carrier that performs material handling and autonomous movement tasks; motion state parameters refer to the real-time speed, acceleration and attitude data of the carrier during operation; suspected obstacle area refers to the three-dimensional spatial range with obstructive morphological features scanned and collected by the environmental detection component; observation baseline refers to the spatial line connecting the reference optical center of the environmental detection component and the geometric center of the suspected obstacle area; displacement target amount refers to the spatial distance measurement required to translate or deflect the observation baseline in order to form an effective observation parallax.

[0052] Specifically, when the obstacle recognition device controls the AGV to perform cruise or standby operations, when the environmental detection component returns the spatial coordinates of a suspected obstacle area, the obstacle recognition device assesses the available safe physical redundancy in the surrounding area based on the current motion state parameters. Based on this safe redundancy and a preset parallax lower limit, it calculates the target displacement that can cause an observable parallax shift in the suspected obstacle area, which serves as the input reference for subsequent physical disturbance strategies.

[0053] In some embodiments, the target displacement of the observation baseline between the environmental detection component and the suspected obstacle area can be calculated in several ways: Optionally, the focal length resolution of the environmental detection component and the initial straight-line distance of the target detection area are extracted, and the required minimum resolution angle is calculated based on the focal length resolution and the initial straight-line distance. The target displacement is then derived using trigonometric functions based on the minimum resolution angle. Optionally, the historical detection pose deviation of the AGV body in the current working coordinate system is retrieved, and a confidence assessment is performed based on the historical detection pose deviation. The calibrated target displacement is then output based on the assessment parameters. It is understood that other constraint-solving methods based on vehicle dynamics models can also be used to generate the target displacement.

[0054] S202. Control the observation baseline to generate a slight disturbance in the displacement target amount relative to the suspected obstacle area, and obtain the detection comparison data of the same suspected obstacle area before and after the slight disturbance.

[0055] Among them, micro-disturbance refers to a short-term displacement operation performed by the control AGV body or environmental detection component relative to the suspected obstacle area, which includes three stages: start-off-reset. The typical duration is 0.2-1.0 seconds, and the total displacement is equal to the target displacement. The detection reference data refers to the multi-dimensional environmental information set, including coordinates, depth, and surface reflection intensity, collected by the environmental detection component for the specific suspected obstacle area at various time nodes before, during, and after the occurrence of the micro-disturbance.

[0056] Specifically, after the displacement target is sent to the execution unit, the obstacle recognition device outputs the corresponding motor drive sequence through the underlying control module. This perturbation physical process is maintained within a preset, configurable short time range and strictly locked at three sampling moments: the initial moment of the perturbation. This is the moment when the motion begins, before the observation baseline has shifted; the moment of maximum disturbance shift. This is the moment when the travel reaches its extreme value, and the observed baseline offset reaches the target displacement; the moment of disturbance recovery. This is the moment when the motion returns to the midpoint, and the observation baseline is reset. The underlying control module of the obstacle recognition device synchronously sends level trigger signals at the above three moments through a hardware timer, instructing the environmental detection component to perform data acquisition, thereby ensuring that the detection and comparison data are strictly aligned with the corresponding spatial pose on the time axis and are encapsulated and backed up. The time data serves as a reference frame to establish the initial coordinates, initial area, and initial echo intensity required for subsequent calculations; The real-time data reflects the maximum parallax response of the target under perturbation excitation and is the main basis for extracting various physical changes; The time-series data is used to verify that the target features can be restored to the baseline state after the disturbance is removed, thereby eliminating misjudgments caused by occasional noise. The above three frames of data together constitute the original input basis for subsequent feature extraction and fingerprint construction.

[0057] In some embodiments, the micro-perturbation of the displacement target amount of the observation baseline relative to the suspected obstacle area and the acquisition of detection control data can be achieved in several ways: Optionally, a predetermined yaw angle is generated by driving the drive unit of the AGV chassis to perform the micro-perturbation, and the camera device is triggered to perform synchronous exposure and point cloud scanning when the steering axis reaches the deflection set value; Optionally, a displacement pulse is sent to the multi-axis translation stage carrying the environmental detection component to make it move relative to the AGV body to achieve the micro-perturbation, and the encoder data of the translation stage is captured synchronously and combined with the captured image frames as control data. It is understood that a multi-sensor joint self-calibration exposure method can also be used to achieve spatiotemporal extraction of the data source.

[0058] S203. Combining the target displacement of the observation baseline, the detection comparison data is inverted to the same viewpoint coordinate system to obtain the parallax displacement and reflection intensity ratio of the suspected obstacle area.

[0059] Among them, the viewpoint coordinate system represents a static three-dimensional spatial reference system established with the reference center of the environmental detection component at the initial moment of the disturbance; the parallax displacement is used to represent the relative offset value of the same surface feature point of the suspected obstacle area in the above reference system after matrix projection alignment; the reflection intensity ratio represents the ratio parameter of microwave echo energy or optical gray value of the same target in multiple time-series data frames.

[0060] Specifically, after aggregating multi-time-series detection and control data, the obstacle recognition device extracts the set of position and attitude transformation parameters recorded during micro-disturbances. Based on this set of parameters, it constructs a spatial inverse transformation matrix and performs inverse projection calculations on the discrete sampling points in the detection and control data. Multi-time-series data points are mapped to the same viewpoint coordinate system to complete clustering and overlay. The system extracts the coordinates of the geometric feature points of this cluster, calculates the Euclidean distance offset of the same feature point due to the influence of the actual physical depth to generate the parallax displacement, and simultaneously analyzes the changes in optical or radio frequency features extracted from various sensor channels to calculate the corresponding reflection intensity ratio.

[0061] In some embodiments, the inversion of detection and control data to the same viewpoint coordinate system and the obtaining of disparity displacement and reflection intensity ratio can be achieved in several ways: Optionally, a spatial voxel mesh model can be constructed using a three-dimensional laser point cloud, the coordinates of local extreme points extracted at each time series can be calculated, and the absolute value of the coordinate difference can be calculated as the disparity displacement after normalizing the extreme points using a quaternion rotation matrix; Optionally, the echo spectrum distribution parameters of the millimeter-wave detector can be extracted, the frequency band energy integral of the maximum deviation time and the initial time in the target clustering region can be calculated, and the energy integrals at the two times can be divided to establish the reflection intensity ratio. It is understood that the inversion and parameter quantization can also be achieved by sensor cross-fusion reprojection.

[0062] S204. Combine the parallax displacement and reflection intensity ratio to calculate the geometric residual and reflection residual, and determine the perturbation fingerprint features of the suspected obstacle area.

[0063] Among them, geometric residual represents the degree of geometric deformation difference between the actual three-dimensional contour of the suspected obstacle region under parallax displacement constraint and the standard rigid solid model; reflection residual refers to the numerical deviation between the actual reflection intensity ratio obtained and the theoretical surface reflection attenuation model; perturbation fingerprint feature is used to represent the high-dimensional feature vector array that identifies the authenticity of the target after the geometric residual and reflection residual are input into a specific mathematical mapping layer and combined in a multi-dimensional normalized manner.

[0064] Specifically, after acquiring the disparity displacement and reflection intensity ratio of the target, the obstacle recognition device quantifies the target's response differences from two physical dimensions. In the geometric dimension, the device compares the relative change rate of the cluster bounding box boundary deformation and cluster area of ​​the target before and after disturbance to assess its contour stability. It also measures the degree of offset divergence of each point cloud node based on the theoretical disparity defined by the observed baseline displacement, thereby calculating the geometric residual. In the reflection dimension, the device evaluates the distribution pattern of the target's texture structure or echo energy sequence under varying viewing angles. It subtracts the actually measured reflection intensity ratio from the ideal diffuse reflection prediction value derived from the viewing angle change to obtain the reflection residual during the matching process with the ideal diffuse reflector. The obstacle recognition device then introduces the upper limit of the range of physical quantities in each dimension as a proportional constant. After dividing the geometric residual and reflection residual by their respective upper limits, the ratio is subtracted from 1 to map back to the 0-1 range. These are then weighted and combined using a preset fusion weighting factor, integrating and compressing into a continuous feature array that is unaffected by changes in the absolute illumination intensity of the scene and can be efficiently computed. This generates a perturbation fingerprint feature with unique characterization properties. After normalization, this feature consists of a first feature quantity representing the significance of lateral offset and a second feature quantity representing contour stability in the geometric dimension, and a third feature quantity representing texture stability and a fourth feature quantity representing echo stability in the reflection dimension. Each feature quantity is stored in fixed-point integer form to adapt to the real-time computing requirements of embedded platforms.

[0065] It should be noted that the data processing for converting basic physical residuals into four-dimensional fingerprint features includes the following: First, the system extracts the maximum bounding box intersection-union ratio (IU / U) of the three-dimensional point cloud before and after micro-perturbation as an evaluation parameter. Subtracting the IU / U from one yields the local contour deformation index. Then, it performs independent weighted matrix calculations with the node displacement deviation obtained by direct inverse normalization of disparity divergence. This refines and decomposes the single macroscopic geometric residual, quantifying it to generate the first feature quantity for characterizing the significance of lateral offset and the second feature quantity for characterizing contour rigidity stability. Similarly, the system extracts the microwave energy deviation returned by the lidar and the contrast change rate of the gray-level co-occurrence matrix captured by the visual sensor. These two reflection response residuals, belonging to different modal hardware channels, are independently back-mapped to the standard data range of 0 to 1 through a logistic regression activation function. This establishes the third feature quantity characterizing surface texture stability and the fourth feature quantity characterizing echo signal stability. Finally, the values ​​of these four independent dimensions are merged to generate a complete perturbation fingerprint feature vector.

[0066] For example, when the system measures that the disparity divergence of a low-lying target is extremely small and the cross-union ratio of its front and rear bounding boxes reaches 0.95, the system automatically converts it into a first feature quantity with a value of 0.92 and a second feature quantity with a value of 0.94 according to the formula. At the same time, if the radar detection echo energy deviation is only 0.04 and the visual surface contrast change rate is 0.03, the system will back-calculate and generate a third feature quantity of 0.95 and a fourth feature quantity of 0.96. These four values ​​representing high physical confidence are packaged to form the four-dimensional perturbation fingerprint of the real entity.

[0067] S205. Calculate the template similarity between the perturbed fingerprint features and the preset entity reference template.

[0068] Among them, the entity reference template represents a multi-dimensional feature benchmark set that is pre-configured in the internal storage space of the system and reflects the standard physical entity under specified micro-perturbation conditions; template similarity refers to the mathematical evaluation parameter of how close the distribution of the currently extracted and generated perturbation fingerprint feature vector is to the entity reference template in the multi-dimensional mapping space.

[0069] Specifically, after constructing the perturbation fingerprint features of the current target, the obstacle recognition device reads the standard feature library in the read-only storage area. This feature library contains at least two types of reference vectors: real entity reference templates. Reference template for false barriers The vector dimensions of both are completely consistent with the perturbation fingerprint features. The values ​​of the real entity reference template in each dimension are close to the upper limit of 1, reflecting the typical characteristics of a real obstacle under perturbation excitation: significant spatial shift, stable contour shape, continuous texture depth, and stable echo intensity. The values ​​of the false obstacle reference template in each dimension are close to the lower limit of 0, reflecting the typical characteristics of a virtual obstacle with light and shadow: no spatial shift, drastic shape changes with angle, disordered texture jumps, and chaotic or missing echoes. The obstacle recognition device uses a spatial distance calculation model to perform difference accumulation evaluation on the extracted high-dimensional feature F and the two types of templates respectively. First, the difference between the current feature and the corresponding dimension of the template is calculated dimension by dimension and squared. Then, the sum of the squares of all dimensions is taken to obtain the Euclidean distance between the current feature and the entity reference template. = and the Euclidean distance between the reference template and the false obstacle .

[0070] The discrimination logic is based on the minimum distance criterion: if Less than This indicates that the current feature is closer to the entity template in the high-dimensional space, and the system determines that the target is a real physical entity; if Less than This indicates that the current feature is closer to a false template, and the system determines that the target is due to ambient light and shadow interference and removes it. Assuming the target is a real physical entity, the system further normalizes the mapping using the reciprocal of the distance to the entity's reference template, calculating the template similarity Sim=1 / (1+ This value quantifies the confidence level of the target's representation of a real object. Obstacle avoidance control commands are only generated when Sim is greater than a preset limit, thus ensuring the accuracy of elimination under the dual discrimination mechanism.

[0071] Taking a specific scenario as an example, let the normalized perturbation fingerprint feature vector be F = [0.90, 0.92, 0.88, 0.85], and the real entity reference template be... =[1.00, 1.00, 1.00, 1.00], False Obstacle Reference Template If the value is [0.00, 0.00, 0.00, 0.00], then the Euclidean distance between the current feature and the entity reference template is the square root of the sum of the squares of the differences in each dimension, i.e.:

[0072] ;

[0073] The Euclidean distance between the current feature and the spurious obstacle reference template is:

[0074] .

[0075] because much smaller Based on the minimum distance criterion, the current feature is closer to the entity template in the high-dimensional feature space, therefore the target is determined to be a real physical entity rather than a false light and shadow interference. Based on this determination, the template similarity Sim=1 / (1+ The value is approximately 0.81, which is greater than the preset limit of 0.60. Based on this, the system confirms that the confidence level of the judgment result is high enough and generates the corresponding obstacle avoidance control command. If we assume that the disturbance fingerprint feature vector of another target is F = [0.05, 0.08, 0.03, 0.06], and the values ​​of each dimension are close to the lower limit, we can calculate... ≈1.89 ≈0.12, because If the target is smaller, the system determines that it is a false obstacle and removes it, preventing it from proceeding to the subsequent similarity assessment stage.

[0076] S206. When the template similarity is greater than the preset limit, determine that there is a real physical entity in the suspected obstacle area and generate the corresponding obstacle avoidance control command.

[0077] Among them, the preset boundary represents the numerical scale used in the business logic to classify the target as an independent obstructing physical entity or a negligible interference source; the real physical entity refers to a three-dimensional structural object that occupies space and has the physical attribute of blocking the AGV vehicle's passage trajectory; the obstacle avoidance control command is used to represent a series of mandatory intervention command messages sent to the motion control terminal when the existence of a threat is confirmed, such as requiring the execution of trajectory deviation, deceleration, or mechanical braking.

[0078] Specifically, the obstacle recognition device receives the template similarity value output by the matching calculation and performs a Boolean comparison with a preset threshold fixed in memory. If the similarity value exceeds the threshold, it indicates that the deformation and reflection of the extracted target object after undergoing slight changes in the viewing angle highly conform to the independent physical laws of the objective entity, eliminating false attributes of ambient light and shadow. Based on this, the device confirms that there is a real physical obstruction at the location, interrupts the original path cruise task queue, refreshes the global spatial information of the object to the impassable layer of the cost map, and sends obstacle avoidance control commands to the underlying execution unit to ensure the physical safety of the vehicle.

[0079] In some embodiments, the determination of real physical entities and the generation of obstacle avoidance control commands can be achieved in several ways: Optionally, after verifying that the template similarity reaches a threshold, a 3D polygonal no-passage marker for the target is generated and added to the local navigation map, while simultaneously triggering a deceleration and emergency stop control pulse from the frequency converter via the control port; Optionally, after determining that the target is a real physical entity, the walkable blank corridor area is calculated based on its occupancy coordinates and contour dimensions, and the parameters of the alternative detour curve are input into a trajectory smoothing generator and encapsulated into a control command packet for transmission to the lower-level machine. It is understood that a multi-stage deceleration braking safety zone triggering method can also be used to issue targeted intervention actions.

[0080] The method provided in this embodiment will now be described in more detail. Please refer to [link / reference]. Figure 3 This is another flowchart illustrating the AGV multi-sensor false obstacle removal method based on vehicle body micro-disturbance fingerprint features in this application embodiment.

[0081] S301. Based on the motion state parameters of the AGV vehicle body and the detected suspected obstacle area, calculate the target displacement of the observation baseline between the environmental detection component and the suspected obstacle area.

[0082] Refer to step S201, which will not be repeated here.

[0083] S302. Calculate the steering wheel deflection amplitude and disturbance duration based on the target displacement and the travel speed in the motion state parameters.

[0084] Among them, the steering wheel deflection amplitude refers to the set angle value by which the steering wheel rotation axis deviates from the straight front of the vehicle; the disturbance duration is used to indicate the length of the data acquisition cycle for maintaining the set angle deflection attitude and performing detection sampling.

[0085] Specifically, after obtaining the desired displacement target, the obstacle recognition device enters the precision motion solution process. Different operating speeds dynamically affect the chassis's following response and the stability of the load center of gravity. The obstacle recognition device calls upon its internally stored vehicle kinematics model, using the driving speed v and the target displacement... Using this as input, calculate the steering wheel deflection amplitude that satisfies the observation baseline offset requirement under dynamic safety constraints. Duration of disturbance The deflection amplitude and duration together determine the actual lateral displacement. Under small-angle approximation conditions, the amount of lateral displacement of the vehicle body... With steering wheel deflection amplitude Driving speed v and duration They satisfy a kinematic relationship ≈v* *sin ,when When the radian value is small, it can be approximated as ≈v* * The system ensures that the actual lateral displacement is not less than the target displacement, i.e. ≥ Thus, the requirement to satisfy parallax is determined. and The feasible region of the parameter pair. Based on this, the system further introduces centrifugal acceleration constraints to prevent the load from tilting due to excessive lateral impulse. The vehicle body performs rotation at a speed v. The lateral acceleration generated at that time is = *tan / L, where L is the wheelbase, system requirements Not exceeding the safe lateral acceleration limit of the load Thus limiting The upper boundary. In Once confirmed, Depend on / (v*sin It can be deduced that the system takes the shortest path while satisfying the above constraints. To minimize the impact on normal cruise operations, the recovery threshold of the deflection angle and the corresponding duration of the disturbance are calculated using kinematic solutions to limit the impact of the operation to strictly within the safety tolerance zone and prevent tilting damage to the transported cargo.

[0086] The calculation process for the above parameters will be illustrated using a specific implementation scenario as an example. Assume the AGV's travel speed is v = 0.5 m / s, and the desired displacement is... =0.05m, upper limit of lateral acceleration for safe load =0.3m / The wheelbase L = 0.6m. First, the theoretical upper limit of the steering wheel deflection amplitude is obtained by solving the centrifugal acceleration constraint:

[0087] =arctan( *L / )=arctan(0.3*0.6 / 0.25)=arctan(0.72)≈35.8 This angle is much larger than the actual required amplitude of the micro-perturbation. While maintaining a safety margin, the system takes... =2 (Approximately 0.035 radians) is taken as the actual deflection amplitude. The corresponding disturbance duration is:

[0088] = / (v*sin =0.05 / (0.5*sin(2) The time taken is approximately 0.05 / (0.5*0.035)≈2.86 seconds, which exceeds the acceptable range for operation because excessive disturbance will significantly affect the normal driving trajectory. In actual engineering implementation, the system will meet the following requirements... Appropriately increase under constraints Up to 3 Or larger, to shorten Alternatively, in the standby state (v≈0), switch to the forward and backward micro-peristaltic mode, directly setting the peristaltic distance equal to... This allows for the dynamic selection of the optimal parameter combination under different operating conditions, achieving efficient parallax excitation while ensuring load safety.

[0089] In some embodiments, the calculation of steering wheel deflection amplitude and disturbance duration can be achieved in several ways: Optionally, a motion mapping table between travel speed and lateral deflection can be queried, and the steering wheel deflection amplitude parameters within the safety limit can be obtained by substituting the target displacement and the safety centrifugal acceleration limit. The corresponding disturbance duration can then be output by dividing the total displacement distance by the calculated lateral velocity. Optionally, a proportional-integral-derivative closed-loop control model based on time-optimal constraints can be constructed, and the current command displacement and travel speed parameters can be input to perform iterative calculations to output a set of steering mechanism control duty cycles and timing durations that optimize both energy consumption and vibration amplitude. It is understood that a control method that uses neural networks for online inference to predict the optimal deflection parameters of the nonlinear suspension chassis can also be employed.

[0090] S303: Combine the steering wheel deflection amplitude and the duration of the disturbance to generate a steering wheel deflection command, so that the AGV body generates a micro-disturbance, and obtain the detection comparison data of the same suspected obstacle area before and after the micro-disturbance.

[0091] Specifically, the obstacle recognition device sends the calculated parameter structure to the servo controller, which converts it into actual steering wheel deflection commands. The motor drive circuit receives the commands and executes on / off switching to generate a tiny yaw angle excitation. Accompanying the vehicle body's extremely short-cycle intervention, the probe assembly is in active listening mode, locking the shutter in batches at the start of the intervention, the moment the movement reaches its extreme value, and the moment of return to the median, collecting physical feedback from the environmental interface, and constructing detection and comparison data with strong temporal alignment characteristics, providing a native physical basis for multi-view calculations.

[0092] Additionally, it should be noted that, regarding the safety issue of potential overturning or resonance caused by directly driving the chassis with slight disturbances under full load and high center of gravity conditions, in some embodiments, to achieve a safer and more flexible disturbance mechanism, the obstacle recognition device acquires the load center of gravity height and wheelbase of the AGV vehicle body, determines the upper limit of lateral acceleration of the AGV vehicle body based on the load center of gravity height and wheelbase, determines the movement speed and stroke of the motion mechanism carrying the environmental detection component based on the upper limit of lateral acceleration and the target displacement, and controls the motion mechanism to move relative to the AGV vehicle body according to the movement speed and stroke, thereby causing slight disturbances to the environmental detection component.

[0093] Among them, the load center of gravity height represents the vertical elevation data of the force node of the overall comprehensive mass of the moving vehicle platform equipped with cargo load from the horizontal driving ground; the wheel track refers to the physical span between the support axles or drive wheels on both sides of the vehicle chassis; the lateral acceleration limit is used to represent the maximum lateral impulse constraint that the upper structure of the vehicle body can withstand under the premise of excluding physical rollover and tire slippage failures; the motion mechanism refers to the electromechanical degree of freedom component installed on the vehicle chassis that is specifically responsible for mounting fixed scanning equipment and can generate rotation or translation operation by relying on an independent drive module.

[0094] Specifically, when the application scenario cannot tolerate chassis vibration or slight steering movements, the obstacle recognition device invokes the center of gravity calculation and counterweight system model, reads the chassis wheelbase geometry to construct a mechanical torsional steady-state defense equation, and derives and outputs the upper limit index of lateral acceleration under this condition to avoid operational safety risks. Based on this safety limit specification, the system then sends an action sequence to the gimbal or sliding rail guide control terminal on the roof, which has independent degrees of freedom. The previously analyzed displacement target is converted into the absolute motion velocity and relative displacement travel of the motion mechanism independently performing sliding deflection. The underlying firmware sends enable and velocity pulses to cause the local components with installed probes to perform relative motion operations on the stationary vehicle body, similarly constructing the generation observation parallax activation feature extraction and obstacle detection links.

[0095] In some embodiments, the physical micro-displacement triggering and process control of external independent components can be achieved in several ways: Optionally, the overall load center of gravity height can be deduced by reading the pressure difference of the base load-bearing module and the count value of the lifting screw, and then the safe lateral acceleration upper limit can be calculated by substituting it into the vehicle body overturning torque model. Using this limit data as the truncated vertex, a smooth trapezoidal acceleration / deceleration curve table can be constructed to drive the external stepping slide rail to perform precise positioning displacement. Optionally, a lookup table matrix of various pre-configured center of gravity configurations and acceleration constraint derating curves can be consulted, and the extracted stroke parameters can be read to configure a high-frequency PWM pulse output to generate a rotational micro-displacement command sequence for the servo turntable equipped with the probe. It is understood that a compliant change in the end-effector posture of a multi-joint industrial robotic arm can also be used to drive the detection probe to complete a six-degree-of-freedom offset intervention operation.

[0096] S304. Combining the target displacement of the observation baseline, the detection and control data are inverted to the same viewpoint coordinate system to obtain the parallax displacement and reflection intensity ratio of the suspected obstacle area.

[0097] Refer to step S203, which will not be repeated here.

[0098] S305. Extract the depth constraint defined by the displacement target of the observation baseline, and determine the geometric residual characterizing the consistency of the three-dimensional structure by combining the disparity displacement.

[0099] Among them, depth constraint refers to the depth limit that objective physical particles in three-dimensional space should maintain, which is pre-calculated based on the change of the observation baseline and the principle of perspective projection; three-dimensional structural consistency refers to the steady-state characteristics of the target group maintaining the surface topology and geometric contour rigid distribution after undergoing spatial observation point transformation.

[0100] Specifically, the obstacle recognition device performs spatial feature calculations on the acquired clusters of detection points with depth information. Based on the generated displacement target parameters, it configures a virtual depth constraint pipeline range containing the theoretical maximum deviation and minimum allowable tolerance. The system maps the 3D point cloud of the suspected obstacle area into this constraint range. The spatial coordinates of the objective entity are highly fixed during inversion and reconstruction, conforming to the characteristics of the pipeline with minimal deviation. Accordingly, the system calculates the degree of distance divergence of each coordinate node in the extracted data relative to the theoretical value of the depth constraint tolerance zone center, and clusters and statistically analyzes the average deformation index of all discrete points in the local area that deviate from the theoretical configuration value, outputting the geometric residual that establishes the consistency of its 3D structure. The depth constraint pipeline is established as follows: According to the principle of perspective projection, when the observation baseline translation distance is... When the straight-line distance from the sensor to the target is D, the theoretical parallax displacement produced by a real entity located on the ground plane in the same viewpoint coordinate system should be: =f* / D, where f is the equivalent focal length parameter of the sensor. Centered on the target, and taking the allowable tolerance δ of the corresponding sensor ranging accuracy, a virtual depth-constrained pipe interval is constructed. −δ, +δ]. For each point cloud node in the suspected obstacle region, calculate its actual disparity after inversion. Compared with theoretical value Deviation | - The arithmetic mean of the deviations of all nodes in the region is used as the geometric residual output. The smaller the value, the higher the consistency of the target's three-dimensional structure and the more it conforms to the physical characteristics of the real entity.

[0101] The calculation process of geometric residuals is illustrated using a specific implementation scenario as an example. Assume the sensor's equivalent focal length f = 500 pixels, and the target displacement... =0.05m, the straight-line distance from the target to the sensor D=2m, then the theoretical parallax =500*0.05 / 2=12.5 pixels. Assume the allowable tolerance δ=1.5 pixels for sensor ranging accuracy, and the depth constraint pipeline is in the [11.0, 14.0] pixel range. If the suspected obstacle area contains N=50 point cloud nodes, in the inversion to... After establishing the time-viewpoint coordinate system, the measured parallax of each node The arithmetic mean is 12.3 pixels, and the deviation of each node is | The arithmetic mean of -12.5| is 0.8 pixels, so the geometric residual output is 0.8 pixels. This residual value is within the allowable tolerance δ=1.5 pixels, indicating that the spatial coordinates of the target are consistent before and after the disturbance, which conforms to the rigid characteristics of the three-dimensional structure of a real entity. If it is a light and shadow illusion, the parallax distribution of its nodes is chaotic and irregular, and the arithmetic mean of the deviation may be as high as 5 pixels or more, far exceeding the tolerance range, resulting in a large geometric residual output, which is then identified as false interference by the subsequent discrimination logic.

[0102] In addition, to address the technical issue that some complex reflective planes not only generate light and shadow but also cause feature confusion due to partial physical occlusion during the subsequent judgment process, in order to improve the accuracy of the elimination logic, in some embodiments, the obstacle recognition device divides the suspected obstacle region into a first component consistent with the depth constraint and a second component inconsistent with the depth constraint based on the sign polarity of the parallax displacement. When the proportion of the second component is greater than a preset proportion threshold, the suspected obstacle region is determined to be a virtual image of the reflective surface and eliminated. When the proportion of the second component is not greater than the preset proportion threshold, the reflection intensity ratio is corrected by the reflection intensity of the first component.

[0103] Among them, the polarity of the sign represents the positive or negative correspondence between the disparity displacement vector obtained by analysis and the direction of the applied physical disturbance; the first component refers to the target information feature point array whose displacement direction conforms to the physical depth distribution constraint; the second component refers to the non-physical interference information feature point array whose displacement exceeds the limit or exhibits an abnormal displacement direction; the preset proportion threshold represents the set proportion limit used to determine whether the overall attribute of the target is biased towards false interference.

[0104] Specifically, when processing the point cloud data sets obtained from the inversion, the obstacle recognition device performs vector segmentation analysis on the disparity displacement of each node in the constituent region. If it is found that the motion polarity of some detection nodes is completely opposite to the set displacement expectation or drastically exceeds the constraint tolerance, this usually indicates that the detection system has captured optical virtual images generated by ground specular reflection or puddle refraction. The device divides compliant coordinate points into the first component memory cluster according to the polarity rule and strips abnormal coordinate points into the second component memory cluster. The proportion parameter of the second component containing abnormal information relative to the total is calculated. If the proportion exceeds the preset proportion threshold, the system directly determines that the main area of ​​the object is a source of interference from the high reflectivity virtual image environment, interrupts and removes subsequent comparison tasks; if the proportion is below the threshold limit, it indicates that it is mainly physical goods covered with a little surface reflection. The system only extracts the echo and intensity data of the first component to correct the parameters of the contaminated reflection intensity ratio data frame and delivers it to the next level of recognition pipeline.

[0105] In some embodiments, component separation and data correction / removal processing can be achieved in various ways: Optionally, the dot product of the disparity vector and the calibration reference vector of each echo scan data point is calculated. If the dot product value is non-negative, it is pushed onto the first component data stack; if it is negative, it is pushed onto the second component data stack. The length data of the two stack areas are read, and the division is performed to obtain the ratio to determine whether to perform a discard operation or to retain the features in the stack and calculate the local average value to overwrite the reflection intensity ratio. Optionally, the image depth gradient operator edge search module is used to crop out the noise image area caused by the drastic change in depth and classify it as the second component. The set of centrally retained pixels is extracted and merged into the first component, and the reflection parameter is recalibrated by calling the regional statistical mean. It is understood that a stripping and reshaping method can also be used to remove sparse boundary interference points and extract the first principal component by three-dimensional spatial density clustering.

[0106] S306. Calculate the mapping function of the reflection intensity ratio as a function of the viewing angle before and after the micro-perturbation, and determine the reflection residual based on the deviation of the reflection intensity ratio and the mapping function.

[0107] Among them, the change in viewing angle represents the angular change in the relative angle between the central axis of the environmental detection light and the normal of the target reflecting surface caused by the application of micro-perturbation; the mapping function is a physical law calculation model that describes the intensity of optical radiation or radio frequency energy return of the irradiated material under the influence of different incident or observation angles; the deviation is used to represent the numerical difference between the actual reflection intensity ratio resolved in multi-frame acquisition and the deduced value given by the mapping function based on physical laws.

[0108] Specifically, the obstacle recognition device analyzes the intensity sequence in video images or radio frequency echoes, extracts the angular fluctuation of the relative surface tangential of the optical sensing center during multi-view acquisition, and establishes the parameter of viewpoint change. The obstacle recognition device extracts... and The change in the angle between the optical center axis and the normal to the target surface at any given time. As a variable representing the change in viewing angle, according to Lambert's law of diffuse reflection, the intensity of reflection from an ideal solid surface is proportional to the cosine of the observation angle, and the system uses this to establish a mapping function. = *cos( + ) / cos( ),in for The initial echo intensity at time t, for The initial observation angle at time [time]. The system imports roughly estimated surface medium properties and a theoretical diffuse reflection function equation library, and derives and calculates the predicted standard reflection intensity ratio that the theoretical environment should feed back under this viewing angle change condition. The actual echo intensity measured at time 1 and Difference and divide by After normalization, we obtain the deviation e = | - | / This refers to the reflection residual. For real objects with stable surface materials, the echo intensity changes with viewing angle according to the aforementioned physical laws, resulting in a small deviation *e*. However, for virtual obstacles or specular reflection sources without a fixed physical depth, their echo intensity fluctuates randomly with the observation angle, leading to a significantly larger deviation *e*, thus effectively distinguishing real targets from environmental interference. The deviation is quantified by subtracting or proportionally calculating the actual measured values ​​from this theoretical prediction, yielding the reflection residual for apparent reflection characteristics. This criterion can efficiently eliminate high-frequency optical noise caused by external illumination.

[0109] The calculation process of reflection residual is illustrated using a specific implementation scenario as an example. Let... Initial observation angle at time =30 Initial echo intensity =800 (the signal-to-noise ratio dimension of a millimeter-wave sensor). After applying a micro-perturbation, The angle between the optical center axis of the ambient detection and the normal to the target surface changes to 32°. Change in perspective Based on the Lambert diffuse reflection mapping function,

[0110] Theoretical predicted intensity = 800 * cos(32) ) / cos(30 = 800 * 0.848 / 0.866 ≈ 784. If the target is a real entity, the reflective properties of its material surface are stable. Real-time measured echo intensity =779, then the reflection residual e=|779−784| / 800=0.00625, which is extremely small, indicating that the target's reflection behavior is highly consistent with physical laws. If the target is a light and shadow illusion, The echo may change or disappear due to angle changes at any given moment; the measured echo intensity =320, then the reflection residual e=|320−784| / 800=0.58. The residual is significantly large, which corresponds to the characteristics of false interference, making it easier for subsequent discrimination logic to remove it.

[0111] In some embodiments, the generation of reflection residuals can be achieved in several ways: Optionally, a local lookup parameter table can be established based on a general model of bidirectional reflection distribution, the theoretical intensity feedback ratio of the target target index can be input as the current analytical viewpoint change, and the root mean square of the absolute difference between the measured data and the index value can be calculated to form a quantified reflection residual; Optionally, a linear decay mathematical equation can be constructed with the cosine component of viewpoint change as the independent variable, the gray-scale measured ratio sequence can be imported into the test parameter distribution deviation index, and the distribution variance or kurtosis bias of the parameter can be obtained as the final reflection residual index. It is understood that a residual establishment mechanism for deriving the prediction error value of optical feature parameters using a pre-trained lightweight neural network can also be adopted.

[0112] S307. Construct perturbation fingerprint features based on geometric residuals and reflection residuals.

[0113] Specifically, the obstacle recognition device aggregates geometric deformation evaluation indicators and surface reflection evaluation indicators extracted from physical responses of different dimensions. A cross-coupled weighted mapping engine is introduced to project the geometric and reflection residuals existing under different physical dimensions onto a pre-defined normalized data space band, merging them to generate a complex matrix array or multi-dimensional index set covering both shape and optical attributes—the so-called perturbation fingerprint feature. This fingerprint extracts the fundamental physical invariants of the objective target unaffected by external perturbations and the corresponding parameters governing the change in viewpoint response, exhibiting excellent information attenuation resistance and fault tolerance robustness.

[0114] In some embodiments, perturbation fingerprint features can be combined and constructed in various ways: Optionally, the geometric residual matrix and the reflection residual sequence are subjected to a linear transformation using a limit-maximum-minimum scaling function, and then sequentially stored in a specific memory area to form a one-dimensional normalized floating-point array data string, which is used to materialize the perturbation fingerprint feature; Optionally, a virtual binary polar coordinate system with the residual values ​​as axes is established, and the polar axis magnitude scale and rotation phase angle of the two residual variables in the reference system plane are calculated to form a complex domain index that can represent the intensity characteristics and category orientation. It is understood that a multivariate principal component analysis algorithm can also be used to remove collinear feature components between data and perform dimensionality reduction and packaging mapping to form a fingerprint descriptor.

[0115] S308. Calculate the template similarity between the perturbed fingerprint features and the preset entity reference template.

[0116] Referring to step S205, it will not be repeated here. It should be noted that template similarity is calculated only after the minimum distance criterion has confirmed that the target is a real physical entity, and is used to quantify the confidence level of the discrimination result. That is, the system first calculates the Euclidean distance between the current perturbed fingerprint feature and the real entity reference template and the false obstacle reference template. and ,like Less than It is initially determined to be a real entity, and then Sim=1 / (1+ ) Calculate template similarity, and only confirm the judgment result and generate obstacle avoidance control instructions if Sim is greater than a preset limit; otherwise Less than If it is identified as a false obstacle, it will be removed without proceeding to the similarity assessment stage. Further details will not be provided here.

[0117] S309. When the template similarity is greater than the preset limit, determine that there is a real physical entity in the suspected obstacle area and generate the corresponding obstacle avoidance control command.

[0118] Refer to step S206, which will not be repeated here.

[0119] In real-world industrial scenarios, in addition to static virtual and physical obstacles, there are also frequent dynamic interference targets such as pedestrians and moving forklifts, which can lead to misjudgments and missed judgments. To accurately isolate dynamic targets from static virtual obstacles, in some embodiments, the obstacle recognition device determines the upper limit of the parallax displacement based on the displacement target amount of the observation baseline and the shortest distance from the suspected obstacle area to the environmental detection component. When the parallax displacement amount is greater than the upper limit, the suspected obstacle area is determined to be a dynamic target, and a corresponding dynamic obstacle avoidance command is generated.

[0120] Among them, the upper limit value represents the maximum theoretical peak value of the parallax that can be calculated by projecting a theoretically stationary physical obstacle under the preset probe relative displacement conditions and detection geometry; dynamic target refers to a three-dimensional coordinate entity object that has a tendency or speed of motion inside and does not have a fixed spatial state; dynamic obstacle avoidance command is used to represent a set of special control command frameworks that are different from static planning algorithms and have velocity prediction characteristics and collision space buffer limit addition attributes.

[0121] Specifically, in the pre-processing stage of inversion matrix and pose analysis, the obstacle recognition device extracts the previously calculated target displacement amount and, together with the length of the three-dimensional distance vector from the device center to the target surface, constructs a virtual observation cone boundary and obtains the values ​​of the vertex values ​​of the possible offset envelope as the upper limit of the disparity displacement amount. The specific calculation method for the upper limit is as follows: based on the proportional scaling relationship of perspective projection, when the sensor performs the target displacement... When the translation and the straight-line distance from the sensor to the target are D, the maximum theoretical parallax produced by a completely stationary target in the viewpoint coordinate system is: =f* / ,in Let f be the closest distance from the target to the sensor, and f be the equivalent focal length. This refers to the theoretical upper limit of the parallax displacement. For a truly stationary target, the measured parallax should not exceed [a certain value] within the measurement error range. If the measured parallax of a suspected obstacle area exceeds This indicates the existence of an independent displacement component relative to the chassis coordinate system within the area, meaning the target possesses autonomous movement attributes and should be classified as a dynamic target. Since the actual displacement coordinates of objective material targets that strictly maintain static attributes are limited by perspective rules and the relative movement amplitude of the detector, if the device detects and analyzes that the displacement of a local point set significantly exceeds or even exceeds this defined upper limit range, this indicates that this spatial feature not only suffers from the parallax of the observation end's movement but is also primarily affected by a bidirectional change in displacement accumulation relative to the chassis coordinate system. It is determined to be personnel or other operational equipment with independent movement dynamics, and then the refined process of identifying virtual and real light and shadow is skipped, directly activating a dynamic obstacle avoidance defense deployment with tracking feedback logic.

[0122] The calculation of the upper limit of parallax and the process of dynamic target recognition are illustrated using a specific implementation scenario as an example. Assume the equivalent focal length of the sensor is f = 500 pixels, and the displacement of the target is... =0.05m, the closest distance from the suspected obstacle area to the sensor. =1.5m, then the theoretical upper limit of the parallax displacement is... =500*0.05 / 1.5≈16.7 pixels. For a completely stationary real entity or a virtual obstacle, the parallax displacement before and after the disturbance should not exceed 16.7 pixels within the measurement error range. If the system detects a measured parallax of 45 pixels in a suspected obstacle area, far exceeding... If the value is 16.7 pixels, it is determined that the target itself has undergone significant displacement during the disturbance, indicating that it possesses independent motion attributes and is identified as a dynamic target. The system skips the virtual / real obstacle fingerprint discrimination process and directly activates the dynamic obstacle avoidance logic, issuing corresponding avoidance commands based on the target's motion trend to ensure vehicle safety.

[0123] In some embodiments, the parallax displacement offset exceeding the limit verification and the corresponding obstacle avoidance command allocation and generation can be achieved in multiple ways: Optionally, the theoretical extreme value tolerance, i.e., the upper limit value, can be calculated using the proportional scaling formula of similar triangles in pinhole imaging and the division quotient of the displacement target amount and the observation distance. If the displacement output exceeds this limit value during multi-frame point cloud matching and iteration, the label is set as a dynamic target, and a tracking controller is introduced to issue dynamic commands. Optionally, the world coordinate system velocity vector of the object can be inferred based on the coordinate transition covariance matrix of different detection channels and spatial mapping nodes. The confidence upper limit of this velocity vector is obtained to judge whether it overflows the static upper limit value and to define the execution of dynamic obstacle avoidance flow. It is understandable that this can also be achieved by calling the underlying Kalman filter prediction model to perform an over-limit threshold check on the differential velocity of the spatial scattered point trajectory.

[0124] After identifying a dynamic target, uniformly issuing a conservative obstacle avoidance strategy would severely reduce workshop traffic efficiency and logistics continuity. To achieve refined control of dynamic game scenarios, in some embodiments, the obstacle recognition device determines the dynamic target's own displacement component based on the parallax displacement and its upper limit; based on the sign polarity of its own displacement component and the displacement direction of the observation baseline, it determines whether the dynamic target is approaching or moving away from the AGV body; when the dynamic target is in the approaching state, a braking command is generated to replace the dynamic obstacle avoidance command; when the dynamic target is in the moving away state, a deceleration observation command is generated to replace the dynamic obstacle avoidance command.

[0125] Among them, the self-displacement component represents the geometric spatial translation data of the target's autonomous movement after removing the excitation displacement component executed by the detector itself from the total value of the global environmental comprehensive parallax displacement data obtained from the solution; the approaching state refers to the three-dimensional physical motion characteristics of the object whose relative velocity vector direction is pointing towards the vehicle's forward travel path and has the potential to cause rigid collision; the moving away state refers to the physical action characteristics of the target object's current movement trajectory gradually deviating from the remote vehicle's planned travel area, thus causing decoupling and separation; the braking command represents the absolute stop guarantee command that forcibly blocks all electric drive axle torque supply and engages the mechanical hard brake mechanism; the deceleration observation command represents the flexible control requirement that does not cut off the main driving logic but only moderately reduces the drive duty cycle to maintain slow following and avoidance while improving the radar warning computing power.

[0126] Specifically, once an object is confirmed to possess independent activity attributes and is classified as a dynamic target, the obstacle recognition device enters the obstacle avoidance level subdivision channel. Using a vector subtraction method, the difference between the obtained comprehensive displacement and the detector's theoretical extreme value is subtracted to extract the target's own displacement component. This displacement component is then compared with the vehicle's existing forward vector using inner product projection and polarity calibration. If the calculated feedback target's relative vector direction forms an acute angle intersection with the vehicle's trajectory, indicating an approaching situation, the avoidance time window is accelerating and decaying. The system determines this as an approaching state and immediately overrides the original control, issuing a braking command to prevent a collision. Conversely, if the target's motion trajectory is parallel and accelerating away or far from the central working projection area, it is determined to be in a far-away state, and no high-intensity emergency braking response is triggered. Instead, a deceleration observation command is issued to maintain the equipment's deceleration and safe following, ensuring unimpeded continuous flow of warehousing and transportation logistics.

[0127] In this embodiment, by employing a technique that uses controllable micro-disturbances of the vehicle body to change the observation baseline to extract parallax displacement and multi-dimensional reflection parameters, and combining geometric residuals and reflection residuals to construct environmentally independent perturbation fingerprint features, the authenticity of the target under test can be quantified from the physical laws of the consistency of the three-dimensional structure and the change of reflection features with the viewing angle. This solves the long-standing technical problem in industrial warehousing where existing passive filtering cannot distinguish between stationary real short objects and stationary light and shadow false obstacles. It achieves high-confidence elimination and safe obstacle avoidance control of multi-sensor false obstacles in complex scenes without increasing hardware overhead.

[0128] The obstacle recognition device in this application embodiment is described below from the perspective of hardware processing. Please refer to [link / reference]. Figure 4 This is a schematic diagram of the physical structure of an obstacle recognition device in the embodiments of this application.

[0129] It should be noted that, Figure 4The structure of the obstacle recognition device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0130] like Figure 4 As shown, the obstacle recognition device includes a CPU 401, which can perform various appropriate actions and processes according to a program stored in ROM 402 or a program loaded into RAM 403 from storage section 408, such as executing the methods in the above embodiments. RAM 403 also stores various programs and data required for system operation. CPU 401, ROM 402, and RAM 403 are interconnected via bus 404. I / O interface 405 is also connected to bus 404.

[0131] The following components are connected to I / O interface 405: input section 406 including audio input devices, push-button switches, etc.; output section 407 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 408 including a hard disk, etc.; and communication section 409 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from them can be installed into storage section 408 as needed.

[0132] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by CPU 401, it performs the various functions defined in this application.

[0133] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0134] Specifically, the obstacle recognition device in this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the AGV multi-sensor false obstacle removal method based on vehicle body micro-disturbance fingerprint features provided in the above embodiment.

[0135] In another aspect, this application also provides a computer-readable storage medium, which may be included in the obstacle recognition device described in the above embodiments; or it may exist independently and not assembled into the obstacle recognition device. The storage medium carries one or more computer programs, which, when executed by a processor of the obstacle recognition device, cause the obstacle recognition device to implement the AGV multi-sensor false obstacle removal method based on vehicle body micro-disturbance fingerprint features provided in the above embodiments.

[0136] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0137] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

Claims

1. A method for eliminating false obstacles in AGVs using multi-sensor sensors based on vehicle body micro-disturbance fingerprint features, characterized in that, The method, applied to an obstacle recognition device, includes: Based on the motion state parameters of the AGV vehicle body and the detected suspected obstacle area, calculate the target displacement of the observation baseline between the environmental detection component and the suspected obstacle area; The observation baseline is controlled to generate a slight perturbation of the displacement target amount relative to the suspected obstacle area, and the detection comparison data of the same suspected obstacle area before and after the slight perturbation are obtained; By combining the displacement target amount of the observation baseline, the detection comparison data is inverted to the same viewpoint coordinate system to obtain the parallax displacement and reflection intensity ratio of the suspected obstacle area; By combining the parallax displacement and the reflection intensity ratio, the geometric residual and reflection residual are calculated to determine the perturbation fingerprint features of the suspected obstacle region; Calculate the template similarity between the perturbation fingerprint features and the preset entity reference template; When the template similarity is greater than a preset limit, it is determined that the suspected obstacle area has a real physical entity, and a corresponding obstacle avoidance control command is generated.

2. The method according to claim 1, characterized in that, The step of calculating the geometric residual and reflection residual by combining the parallax displacement and the reflection intensity ratio to determine the perturbation fingerprint features of the suspected obstacle region specifically includes: The depth constraint defined by the displacement target of the observation baseline is extracted, and the geometric residual characterizing the consistency of the three-dimensional structure is determined by combining the disparity displacement. The reflection intensity ratio is calculated as a function of the viewing angle before and after the micro-perturbation, and the reflection residual is determined based on the deviation between the reflection intensity ratio and the mapping function. Perturbation fingerprint features are constructed based on the geometric residual and the reflection residual.

3. The method according to claim 2, characterized in that, After the step of extracting the depth constraint defined by the displacement target amount of the observation baseline and determining the geometric residual characterizing the consistency of the three-dimensional structure in combination with the disparity displacement amount, the method further includes: Based on the sign polarity of the disparity displacement, the suspected obstacle region is divided into a first component consistent with the depth constraint and a second component inconsistent with the depth constraint; When the proportion of the second component is greater than a preset proportion threshold, the suspected obstacle area is determined to be a virtual image of the reflective surface and is removed. When the proportion of the second component is not greater than the preset proportion threshold, the reflection intensity ratio is corrected by the reflection intensity of the first component.

4. The method according to claim 1, characterized in that, The step of controlling the observation baseline to generate the micro-perturbation of the displacement target amount relative to the suspected obstacle region specifically includes: Based on the target displacement and the travel speed in the motion state parameters, calculate the steering wheel deflection amplitude and the duration of the disturbance. The steering wheel deflection amplitude and the duration of the disturbance are combined to generate a steering wheel deflection command, which causes a slight disturbance to the AGV body.

5. The method according to claim 1, characterized in that, The step of controlling the observation baseline to generate the micro-perturbation of the displacement target amount relative to the suspected obstacle region specifically includes: Obtain the load center of gravity height and wheelbase of the AGV vehicle body, and determine the upper limit of lateral acceleration of the AGV vehicle body based on the load center of gravity height and wheelbase; Based on the upper limit of lateral acceleration and the target displacement, determine the motion speed and stroke of the motion mechanism carrying the environmental detection component; The motion mechanism is controlled to move relative to the AGV body at the specified speed and travel distance, causing the environmental detection component to generate slight disturbances.

6. The method according to claim 1, characterized in that, Before the step of calculating the template similarity between the perturbed fingerprint features and the preset entity reference template, the method further includes: The upper limit of the parallax displacement is determined based on the target displacement of the observation baseline and the nearest distance from the suspected obstacle area to the environmental detection component; When the parallax displacement is greater than the upper limit, the suspected obstacle area is determined to be a dynamic target, and a corresponding dynamic obstacle avoidance command is generated.

7. The method according to claim 6, characterized in that, After determining the suspected obstacle area as a dynamic target and generating a corresponding dynamic obstacle avoidance command when the parallax displacement is greater than the upper limit value, the method further includes: The self-displacement component of the dynamic target is determined based on the disparity displacement and the upper limit value; Based on the sign polarity of its own displacement component and the displacement direction of the observation baseline, the approaching or moving away state of the dynamic target relative to the AGV body is determined; When the dynamic target is in the approaching state, a braking command is generated to replace the dynamic obstacle avoidance command; When the dynamic target is in the distance state, a deceleration observation command is generated to replace the dynamic obstacle avoidance command.

8. An obstacle recognition device, characterized in that, The obstacle recognition device includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the obstacle recognition device to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the obstacle recognition device, the obstacle recognition device performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the obstacle recognition device, the obstacle recognition device performs the method as described in any one of claims 1-7.

Images