Vehicle control method, vehicle, and computer-readable storage medium

Abstract

The embodiment of the application provides a vehicle control method, a vehicle and a computer readable storage medium, the method comprises the following steps: obtaining an observation state sequence and a preset state sequence; constructing a composite distance function based on the observation state sequence and the preset state sequence; comparing the similarity distance with the preset difference threshold to obtain a comparison result; and in response to the comparison result and the termination coordinate position satisfying a preset parking condition, performing parking control on the target vehicle. The application solves the technical problems of low recognition accuracy and low use safety of the related gesture recognition vehicle parking-out system.

Description

Technical Field

[0001] This application relates to the field of vehicle technology, and more specifically, to a vehicle control method, a vehicle, and a computer-readable storage medium. Background Technology

[0002] With the development of intelligent driving assistance systems, automatic parking functions have been widely applied in mass-produced vehicles, and gesture interaction, as a natural human-machine control method, is gradually being used in parking scenarios. Related technologies primarily rely on a monocular camera to capture user gestures and identify parking commands, such as circling or waving a hand, through two-dimensional image trajectory matching. However, these solutions are prone to missed or false detections due to loss of visual information in bright light, backlight, or nighttime environments. Furthermore, the information dimension is limited, relying solely on hand paths, which can lead to a limited variety of commands and confusion. In addition, these technologies lack spatial safety verification, failing to distinguish between legitimate close-range user operations and accidental touches from a distance, posing safety risks.

[0003] There is currently no good solution to the above problems. Summary of the Invention

[0004] This application provides a vehicle control method, a vehicle, and a computer-readable storage medium to at least solve the technical problems of low recognition accuracy and low security in related gesture recognition vehicle parking systems.

[0005] According to one aspect of the embodiments of this application, a vehicle control method is provided, comprising: acquiring an observed state sequence and a preset state sequence, wherein the observed state sequence is used to represent the real-time operation change state corresponding to a target part of a target user within a preset time window, the observed state sequence including the termination coordinate position corresponding to the target part, and the preset state sequence is used to represent the standard operation change state corresponding to a preset parking intention; constructing a composite distance function based on the observed state sequence and the preset state sequence, wherein the composite distance function is used to determine the similarity distance between the observed state sequence and the preset state sequence; comparing the similarity distance with a preset difference threshold to obtain a comparison result, wherein the comparison result is used to determine the target parking intention corresponding to the observed state sequence; and performing parking control on the target vehicle in response to the comparison result and the termination coordinate position satisfying a preset parking condition.

[0006] Optionally, obtaining the observation state sequence includes: in response to the distance between the target user and the target vehicle being less than or equal to a preset distance threshold, acquiring visual perception data and radar perception data corresponding to the target user, wherein the visual perception data and radar perception data are obtained by data collection of the target user through an onboard multi-source sensing unit; detecting the visual perception data based on a target detection model to obtain an initial detection box, wherein the initial detection box is used to represent the two-dimensional bounding box corresponding to the target part, the target detection model is obtained by training on sample data and the labels corresponding to the sample data, the sample data is used to represent historical visual perception data, and the labels corresponding to the sample data are used to represent the detection boxes corresponding to the historical visual perception data; performing cluster analysis on the radar perception data to obtain multiple clusters, wherein the multiple clusters are used to determine the dynamic targets in the current environment of the target user; and determining the observation state sequence based on the initial detection box and the multiple clusters.

[0007] Optionally, determining the observation state sequence based on the initial detection box and multiple clusters includes: transforming the initial detection box based on preset calibration parameters to obtain an initial search area, wherein the initial search area is used to represent the three-dimensional search area of ​​the target part in the radar coordinate system, wherein the preset calibration parameters include a preset intrinsic parameter matrix and a preset extrinsic parameter matrix; and determining the observation state sequence based on the initial detection box, the initial search area, and multiple clusters.

[0008] Optionally, determining the observation state sequence based on the initial detection box, the initial search region, and multiple clusters includes: determining the target cluster among the multiple clusters based on the initial search region and multiple clusters, wherein the target cluster is used to represent the dynamic point cloud cluster corresponding to the target part; and determining the observation state sequence based on the target cluster and the initial detection box.

[0009] Optionally, determining the target cluster among the multiple clusters based on the initial search region and multiple clusters includes: obtaining the centroid and radial velocity of the current cluster among the multiple clusters; and determining the current cluster as the target cluster in response to the fact that the centroid of the current cluster is located in the initial search region and the radial velocity satisfies a preset feature condition.

[0010] Optionally, the radar sensing data further includes: the three-dimensional instantaneous velocity corresponding to the target location; determining the observation state sequence based on the target cluster and the initial detection box includes: determining the target detection box based on the target cluster and the initial detection box, wherein the target detection box is used to determine the spatial coordinate position corresponding to the target location; extracting features from the target detection box to obtain the target location contour; fitting an ellipse to the target location contour to obtain a target fitting function, wherein the target fitting function is used to determine the pointing angle of the target location contour within the target detection box; and determining the observation state sequence within a preset time window based on the target detection box, the target fitting function, and the three-dimensional instantaneous velocity corresponding to the target location.

[0011] Optionally, the preset state sequence includes a first state sequence and a second state sequence, with the parking directions corresponding to the first state sequence and the second state sequence being opposite. In response to the comparison result and the termination coordinate position satisfying preset parking conditions, parking control of the target vehicle includes: in response to the similarity distance between the observed state sequence and the first state sequence being less than or equal to a preset difference threshold, and the termination coordinate position being located within a preset safety trigger area, controlling the target vehicle to perform a parking operation based on a first direction, wherein the first direction is the parking direction corresponding to the first state sequence; or, in response to the similarity distance between the observed state sequence and the second state sequence being less than or equal to a preset difference threshold, and the termination coordinate position being located within a preset safety trigger area, controlling the target vehicle to perform a parking operation based on a second direction, wherein the second direction is the parking direction corresponding to the second state sequence.

[0012] Optionally, the vehicle control method further includes: determining that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the similarity distance between the observed state sequence and the first state sequence being greater than a preset difference threshold; or, determining that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the similarity distance between the observed state sequence and the second state sequence being greater than a preset difference threshold; or, determining that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the termination coordinate position not being located in a preset safety triggering area.

[0013] According to another aspect of the embodiments of this application, a vehicle control device is also provided, comprising: an acquisition module, configured to acquire an observation state sequence and a preset state sequence, wherein the observation state sequence represents the real-time operation change state corresponding to a target part of a target user within a preset time window, the observation state sequence includes the termination coordinate position corresponding to the target part, and the preset state sequence represents the standard operation change state corresponding to a preset parking intention; a construction module, configured to construct a composite distance function based on the observation state sequence and the preset state sequence, wherein the composite distance function is used to determine the similarity distance between the observation state sequence and the preset state sequence; a comparison module, configured to compare the similarity distance with a preset difference threshold to obtain a comparison result, wherein the comparison result is used to determine the target parking intention corresponding to the observation state sequence; and a control module, configured to perform parking control on the target vehicle in response to the comparison result and the termination coordinate position satisfying a preset parking condition.

[0014] Optionally, the acquisition module is further configured to: in response to the distance between the target user and the target vehicle being less than or equal to a preset distance threshold, acquire visual perception data and radar perception data corresponding to the target user, wherein the visual perception data and radar perception data are obtained by data collection of the target user through an onboard multi-source sensing unit; detect the visual perception data based on a target detection model to obtain an initial detection box, wherein the initial detection box is used to represent the two-dimensional bounding box corresponding to the target part, the target detection model is obtained by training on sample data and the labels corresponding to the sample data, the sample data is used to represent historical visual perception data, and the labels corresponding to the sample data are used to represent the detection boxes corresponding to the historical visual perception data; perform cluster analysis on the radar perception data to obtain multiple clusters, wherein the multiple clusters are used to determine the dynamic targets in the current environment of the target user; and determine the observation state sequence based on the initial detection box and the multiple clusters.

[0015] Optionally, the acquisition module is further configured to: transform the initial detection box based on preset calibration parameters to obtain an initial search region, wherein the initial search region is used to represent the three-dimensional search region of the target part in the radar coordinate system, wherein the preset calibration parameters include a preset intrinsic parameter matrix and a preset extrinsic parameter matrix; and determine the observation state sequence based on the initial detection box, the initial search region and multiple clusters.

[0016] Optionally, the acquisition module is further configured to: determine a target cluster among multiple clusters based on an initial search region and multiple clusters, wherein the target cluster is used to represent the dynamic point cloud cluster corresponding to the target location; and determine an observation state sequence based on the target cluster and the initial detection box.

[0017] Optionally, the acquisition module is further configured to: acquire the centroid and radial velocity of the current cluster among multiple clusters; and determine the current cluster as the target cluster in response to the fact that the centroid of the current cluster is located in the initial search region and the radial velocity satisfies the preset feature conditions.

[0018] Optionally, the radar sensing data also includes: the three-dimensional instantaneous velocity corresponding to the target part. The acquisition module is further used to: determine the target detection box based on the target cluster and the initial detection box, wherein the target detection box is used to determine the spatial coordinate position corresponding to the target part; extract features from the target detection box to obtain the target part contour; perform ellipse fitting on the target part contour to obtain the target fitting function, wherein the target fitting function is used to determine the pointing angle of the target part contour in the target detection box; and determine the observation state sequence within a preset time window based on the target detection box, the target fitting function, and the three-dimensional instantaneous velocity corresponding to the target part.

[0019] Optionally, the preset state sequence includes a first state sequence and a second state sequence, wherein the parking driving directions corresponding to the first state sequence and the second state sequence are opposite. The control module is further configured to: control the target vehicle to perform a parking operation based on a first direction in response to the similarity distance between the observed state sequence and the first state sequence being less than or equal to a preset difference threshold, and the termination coordinate position being located in a preset safety trigger area; or, control the target vehicle to perform a parking operation based on a second direction in response to the similarity distance between the observed state sequence and the second state sequence being less than or equal to a preset difference threshold, and the termination coordinate position being located in a preset safety trigger area, wherein the second direction is the parking driving direction corresponding to the second state sequence.

[0020] Optionally, the vehicle control device further includes: a determination module, configured to: determine that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the similarity distance between the observed state sequence and the first state sequence being greater than a preset difference threshold; or, determine that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the similarity distance between the observed state sequence and the second state sequence being greater than a preset difference threshold; or, determine that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the termination coordinate position not being located in a preset safety triggering area.

[0021] According to another aspect of the embodiments of this application, a vehicle is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0022] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0023] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0024] According to another aspect of the embodiments of this application, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods in various embodiments of this application.

[0025] According to another aspect of the embodiments of this application, a computer program is also provided, which, when executed by a processor, implements the methods of the various embodiments of this application.

[0026] In this embodiment, a combination of multi-sensor fusion and geometric feature analysis is employed. This involves acquiring an observed state sequence and a preset state sequence. The observed state sequence represents the real-time operational changes of a target user's target location within a preset time window, including the termination coordinates of the target location. The preset state sequence represents the standard operational changes corresponding to a preset parking intent. A composite distance function is then constructed based on the observed and preset state sequences to determine the similarity distance between them. This similarity distance is then compared to a preset difference threshold to obtain a comparison result. This comparison result determines the target parking intent corresponding to the observed state sequence. Finally, in response to the comparison result and termination coordinates satisfying preset parking conditions, parking control is applied to the target vehicle. This achieves accurate identification of legitimate user intent and eliminates environmental interference and misoperation, thus realizing a high-accuracy, high-security, and highly interactive gesture-controlled parking effect. This solves the technical problems of low recognition accuracy and low safety in related gesture recognition vehicle parking systems. Attached Figure Description

[0027] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0028] Figure 1 This is a flowchart of an optional vehicle control method according to an embodiment of this application;

[0029] Figure 2 This is a schematic diagram illustrating an optional data acquisition method according to an embodiment of this application;

[0030] Figure 3 This is a schematic diagram of an optional feature extraction method according to an embodiment of this application;

[0031] Figure 4 This is a schematic diagram of an optional parking control according to an embodiment of this application;

[0032] Figure 5 This is a structural block diagram of an optional vehicle control device according to an embodiment of this application. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0034] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0035] With the rapid development of autonomous driving assistance technology, automatic parking systems, such as Remote Parking Assist (RPA) and Memory Parking Assist (HPA), have become standard features in mid-to-high-end vehicles. To enhance the naturalness and convenience of human-vehicle interaction, gesture recognition-based parking control technology has gradually emerged. Most related solutions use a monocular camera to capture user gesture images and then match hand movement trajectories, such as circling or waving, using preset templates to trigger automatic parking commands. However, these technologies have significant drawbacks: First, they heavily rely on visual information. In strong light, backlight, low illumination, or complex backgrounds, such as underground parking garage walls and parking lines, hand detection is prone to failure, leading to missed detections or false triggers. Second, these technologies only extract two-dimensional trajectory information, ignoring key posture features such as palm orientation and finger spread. The types of commands that can be recognized are limited, and similar actions, such as waving and pointing, are difficult to distinguish. Third, these solutions lack awareness of the relative three-dimensional position of the user and the vehicle, making it impossible to determine whether the operation is within a safe and effective distance. For example, within 5 meters, it is easily interfered with by unauthorized personnel from a distance. Fourth, the relevant recognition algorithms typically use fixed thresholds and do not consider individual differences in user body size, gesture range, or movement speed, resulting in poor adaptability and inconsistent user experience. These issues collectively limit the practicality and security of the gesture-based parking function.

[0036] According to an embodiment of this application, a method embodiment for vehicle control is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0037] This embodiment provides a vehicle control method. Figure 1 This is a flowchart of an optional vehicle control method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps:

[0038] Step S102: Obtain the observation state sequence and the preset state sequence. The observation state sequence is used to represent the real-time operation change state corresponding to the target part of the target user within the preset time window. The observation state sequence includes the termination coordinate position corresponding to the target part. The preset state sequence is used to represent the standard operation change state corresponding to the preset parking intention.

[0039] Step S104: Based on the observed state sequence and the preset state sequence, construct a composite distance function, wherein the composite distance function is used to determine the similarity distance between the observed state sequence and the preset state sequence;

[0040] Step S106: Compare the similarity distance and the preset difference threshold to obtain the comparison result, wherein the comparison result is used to determine the target parking intention corresponding to the observed state sequence;

[0041] Step S108: In response to the comparison result and the termination coordinate position satisfying the preset parking conditions, parking control is performed on the target vehicle.

[0042] The aforementioned target body part refers to the human body part continuously tracked and identified by the multi-sensor system during the user's parking operation, used to represent the intention of the operation. This can be the palm, hand area, lower forearm, fingertips, the center of the wrist in a clenched fist state, or the wrist area when wearing an interactive smart bracelet, etc. This application uses the palm as an example to illustrate the following related embodiments.

[0043] The above observation state sequence represents a set of dynamic change state data of the target user's target location in three-dimensional space, continuously collected by a multi-sensor fusion system within a preset time window. It includes the three-dimensional spatial coordinates of the target location at each moment. The sequence includes the three-dimensional instantaneous velocity, palm rotation angle, and angular velocity change, and the terminal element of the sequence contains the three-dimensional termination coordinates of the target part at the moment the gesture is completed. ), used to fully characterize the spatiotemporal evolution of a gesture from beginning to end.

[0044] The aforementioned preset state sequence represents a pre-stored sequence of standard operation change state data corresponding to a specific parking intention, such as parking out on the left or right. The preset state sequence consists of the three-dimensional coordinates, speed, rotation angle, and angular velocity changes collected within the same time window for the standardized hand gesture, serving as a reference benchmark for intention matching.

[0045] The aforementioned composite distance function represents a mathematical function used to quantify the similarity between the observed state sequence and the preset state sequence. It is composed of a weighted sum of the L2 norm difference of the target part's three-dimensional velocity sequence, the absolute value deviation of the hand rotation angle, and the L1 norm difference of the Hu invariant moments sequence, weighted according to calibration weights. This function comprehensively reflects the overall similarity of the gesture in terms of trajectory, posture, and geometric features. Specifically, the L2 norm difference of the target part's three-dimensional velocity sequence measures the consistency of movement speed. The absolute value deviation of the hand rotation angle assesses whether the gesture direction aligns with the standard template. The L1 norm difference of the Hu invariant moments sequence describes the geometric stability of the hand contour shape under rotation and translation.

[0046] The aforementioned preset difference threshold represents the system's preset similarity judgment threshold, used to distinguish between valid gestures and invalid interference actions. When the similarity distance calculated by the composite distance function is lower than this threshold, it is determined that the observed state sequence successfully matches a preset state sequence, thereby identifying the corresponding target parking intention.

[0047] The above comparison results represent the judgment output generated by comparing the similarity distance with a preset difference threshold. The output is either "match successful" or "match failed," indicating whether the observed state sequence matches the parking intention corresponding to a preset state sequence. Figure 1 To.

[0048] The above-mentioned preset parking conditions indicate that the comparison result is a successful match, and the three-dimensional termination coordinates of the target part are located within the preset safe trigger area. The preset safe trigger area can be within 2.5 meters in front of or behind the vehicle and 0.8 to 1.8 meters in height.

[0049] In the process of gesture recognition, this application first synchronously acquires real-time dynamic data of the current user's actions, i.e., the observation state sequence, and a predefined legal parking action template, i.e., the preset state sequence. The observation state sequence includes not only the movement trajectory and posture changes of the gesture, but also the precise spatial termination position when the gesture ends, so as to achieve the synchronous expression of action integrity and spatial compliance. The preset state sequence provides a standardized reference model for each parking intention, enabling the system to accurately match real-time actions with preset intentions through structural alignment.

[0050] After acquiring the observed state sequence and the preset state sequence, this embodiment of the application constructs a composite distance function based on the observed state sequence and the preset state sequence. That is, the system designs a comprehensive mathematical function based on the multi-dimensional state data that are structurally completely consistent with the observed state sequence and the preset state sequence, including but not limited to the three-dimensional spatial coordinates of the target part, the three-dimensional instantaneous velocity, the palm rotation angle and the change in angular velocity, to quantify the overall degree of difference between the current user's gesture action and the preset parking intention template, that is, the similarity distance.

[0051] After obtaining the similarity distance, this embodiment of the application further compares the similarity distance with a preset difference threshold. The system compares the calculated similarity distance with the preset difference threshold and performs the following logical judgment operation: if the similarity distance is less than or equal to the preset difference threshold, it is determined that the observed state sequence matches the corresponding preset state sequence successfully; if the similarity distance is greater than the preset difference threshold, it is determined that the match fails. The generated comparison result uniquely corresponds to a certain preset parking intention. That is, when the comparison result is "match successful", the system binds the result with the parking intention associated with the compared preset state sequence, thereby clarifying the specific parking operation intention represented by the current gesture, such as recognizing it as a "parking on the right" or "parking on the left" instruction.

[0052] Then, this embodiment of the application determines whether the comparison result and the termination coordinate position meet the preset parking conditions in order to perform parking control on the target vehicle. The system only enters the parking control activation state when the comparison result is "successfully matched" and the termination coordinate position is within the safe trigger area; if either condition is not met, the system will not initiate control commands. After the preset parking conditions are met, the parking domain controller sends preset control signals to the steering actuator, drive motor and braking system of the target vehicle through the Controller Area Network (CAN) bus to start the automatic parking program and control the vehicle to complete the reverse driving action from the perpendicular parking space along the preset path.

[0053] Based on steps S102 to S108, this embodiment of the application adopts a combination of multi-sensor fusion and geometric feature analysis. It acquires an observation state sequence and a preset state sequence. The observation state sequence represents the real-time operational change state corresponding to the target part of the target user within a preset time window, including the termination coordinate position corresponding to the target part. The preset state sequence represents the standard operational change state corresponding to the preset parking intention. Then, based on the observation state sequence and the preset state sequence, a composite distance function is constructed to determine the similarity distance between the observation state sequence and the preset state sequence. The similarity distance is then compared with a preset difference threshold to obtain a comparison result. This comparison result is used to determine the target parking intention corresponding to the observation state sequence. Finally, in response to the comparison result and the termination coordinate position satisfying the preset parking conditions, parking control is applied to the target vehicle. This achieves the goal of accurately identifying legitimate user intentions and eliminating environmental interference and misoperation, thereby realizing the technical effect of high accuracy, high security, and high interactive richness of gesture-controlled parking. This solves the technical problems of low recognition accuracy and low safety in related gesture recognition vehicle parking systems.

[0054] Optionally, obtaining the observation state sequence includes: in response to the distance between the target user and the target vehicle being less than or equal to a preset distance threshold, acquiring visual perception data and radar perception data corresponding to the target user, wherein the visual perception data and radar perception data are obtained by data collection of the target user through an onboard multi-source sensing unit; detecting the visual perception data based on a target detection model to obtain an initial detection box, wherein the initial detection box is used to represent the two-dimensional bounding box corresponding to the target part, the target detection model is obtained by training on sample data and the labels corresponding to the sample data, the sample data is used to represent historical visual perception data, and the labels corresponding to the sample data are used to represent the detection boxes corresponding to the historical visual perception data; performing cluster analysis on the radar perception data to obtain multiple clusters, wherein the multiple clusters are used to determine the dynamic targets in the current environment of the target user; and determining the observation state sequence based on the initial detection box and the multiple clusters.

[0055] The aforementioned preset distance threshold represents the minimum safe distance preset by the system to determine whether a target user has entered the gesture recognition activation range. Its value can be 5 meters to ensure that the sensing and acquisition process is only initiated when a legitimate user approaches the vehicle, thus avoiding false wake-up in unauthorized environments.

[0056] The aforementioned visual perception data represents a sequence of image frames collected by an onboard visual sensor, containing two-dimensional pixel information of the target user and their surrounding environment. The data source is a camera module deployed around the vehicle.

[0057] The radar perception data mentioned above refers to the raw point cloud data collected by the vehicle-mounted millimeter-wave radar, which includes the three-dimensional coordinates, radial velocity and reflection intensity information of dynamic targets, and is used to identify human limb movements.

[0058] The aforementioned vehicle-mounted multi-source sensing unit can be a combined hardware system of visual sensors and millimeter-wave radar deployed around the target vehicle. Their collaborative operation enables spatial perception and motion capture of the target user.

[0059] The aforementioned object detection model represents a lightweight convolutional neural network based on deep learning. It undergoes supervised training using a large amount of historical visual perception data and its corresponding labeled detection boxes, enabling the model to accurately output 2D bounding boxes of target locations in complex environments such as strong light, backlight, nighttime, and cluttered backgrounds. The object detection model can be YOLOv5 (You Only Look Once version 5 small, YOLOv5s), MobileNetV3 (MobileNetV3+Single Shot MultiBox Detector, MobileNetV3-SSD), etc. The specific object detection model is not limited here.

[0060] The initial detection box described above represents the rectangular region enclosed by the target detection model in the image plane, representing the target area. Its coordinate form is: This is used to define the region for visual semantic recognition. These represent the minimum pixel coordinates of the bounding box in the horizontal direction (X-axis) of the image, i.e., the column index of the left edge of the bounding box; the minimum pixel coordinates of the bounding box in the vertical direction (Y-axis) of the image, i.e., the row index of the top edge of the bounding box; the maximum pixel coordinates of the bounding box in the horizontal direction (X-axis) of the image, i.e., the column index of the right edge of the bounding box; and the maximum pixel coordinates of the bounding box in the vertical direction (Y-axis) of the image, i.e., the row index of the bottom edge of the bounding box.

[0061] The sample data mentioned above represents historical visual perception images from real-world scenarios used during the training phase. These images cover different lighting conditions, user body sizes, gestures, and background environments, and are used to build the model's generalization ability and ensure its adaptability in real-world in-vehicle scenarios.

[0062] The labels corresponding to the above sample data represent the coordinates of the target part detection box that are manually or semi-automatically labeled and precisely correspond to each sample data. They are used to provide supervision signals for the model and guide it to learn the visual morphology and spatial distribution patterns of the target parts.

[0063] The above clustering analysis indicates that a density-based clustering algorithm (Density-Based Spatial Clustering of Applications with Noise, DBSCAN) is used on radar sensing data to filter out points in the point cloud that belong to the static background, retaining only the dynamic target point cloud with low-speed movement characteristics, forming multiple independent clusters, and each cluster represents a potential moving target.

[0064] Before acquiring visual and radar perception data corresponding to the target user, the system must ensure that the distance between the target user and the target vehicle is less than or equal to a preset distance threshold. That is, the gesture recognition process is only activated when the system detects that the distance between the target user and the vehicle does not exceed the preset distance threshold via smart key signal or near-field communication. This ensures that the system only initiates sensor acquisition when the authorized user enters the effective interaction range, avoiding continuous operation in non-target scenarios and reducing power consumption and the risk of false triggering. After the trigger condition is met, the system simultaneously activates the cameras and millimeter-wave radar deployed around the vehicle to acquire image sequences and point cloud data containing the user and their environment. Visual perception data provides the appearance semantics of the target area, while radar perception data provides its three-dimensional spatial position and movement speed; the two constitute complementary multimodal perception inputs.

[0065] After acquiring visual perception data, the system uses an object detection model to detect the data and obtain initial detection boxes. Specifically, the system inputs the acquired image frames into an object detection model composed of a lightweight convolutional neural network. This model performs semantic analysis on the image and outputs a two-dimensional rectangular region to locate the precise position of the target part in the image plane.

[0066] After acquiring radar sensing data, cluster analysis is performed to obtain multiple clusters. Specifically, the system uses the DBSCAN density clustering algorithm to process millimeter-wave radar point cloud data, removing stationary background points such as walls, ground, and parking lines, while retaining dynamic point groups with low-speed movement characteristics, forming several independent clusters. Each cluster represents a potential moving target. The spatial distribution and motion characteristics of multiple clusters are used to distinguish effective interactive targets such as the user's hand from other moving interference objects in the environment, such as pedestrians and vehicles, providing a set of candidate targets for accurate correlation between visual and radar data.

[0067] When determining the observation state sequence based on the initial detection box and multiple clusters, the system associates and matches the two-dimensional position information of the target part obtained by visual perception with the three-dimensional spatial information of the dynamic target obtained by radar perception, thereby comprehensively obtaining a complete data sequence representing the motion state of the target part in three-dimensional space. This sequence covers the spatial coordinates, velocity and attitude change characteristics of the target part.

[0068] Based on the above optional embodiments, this application embodiment achieves on-demand activation of the perception system by setting a preset distance threshold, reducing invalid computation and power consumption. A target detection model trained with sample data and corresponding labels is used to achieve high-precision two-dimensional positioning of target parts in complex visual environments. Dynamic targets are separated from radar perception data through DBSCAN clustering analysis, effectively suppressing background interference. High-reliability fusion of visual semantics and radar ranging is achieved through cross-modal spatial matching of the initial detection box and multiple clusters, ultimately generating an observation state sequence with three-dimensional spatiotemporal integrity, significantly improving the robustness, accuracy, and environmental adaptability of gesture recognition.

[0069] Figure 2 This is a schematic diagram illustrating an optional data acquisition method according to an embodiment of this application, such as... Figure 2 As shown, during automatic parking control, the system initially enters a low-power sleep state. When the smart key detects a legitimate user entering within 5 meters, the parking domain controller is awakened, simultaneously activating sensors such as the surround-view camera, 77GHz millimeter-wave radar, and ultrasonic sensor array. The vision unit outputs a real-time two-dimensional pixel bounding box of the hand using a lightweight neural network, while the radar unit preprocesses the point cloud, including background filtering and DBSCAN clustering, preserving dynamic targets—that is, extracting dynamic target clusters—to obtain relevant two-dimensional coordinates.

[0070] Optionally, determining the observation state sequence based on the initial detection box and multiple clusters includes: transforming the initial detection box based on preset calibration parameters to obtain an initial search area, wherein the initial search area is used to represent the three-dimensional search area of ​​the target part in the radar coordinate system, wherein the preset calibration parameters include a preset intrinsic parameter matrix and a preset extrinsic parameter matrix; and determining the observation state sequence based on the initial detection box, the initial search area, and multiple clusters.

[0071] The aforementioned preset calibration parameters represent a set of fixed parameters obtained through an offline calibration process before system deployment, describing the geometric and coordinate relationships between the vision sensor and the radar sensor. These parameters include a preset intrinsic parameter matrix and a preset extrinsic parameter matrix. The preset intrinsic parameter matrix is ​​a pre-calibrated camera intrinsic parameter matrix K, used to describe the camera's optical imaging characteristics. The preset extrinsic parameter matrix is ​​the radar-camera extrinsic parameter matrix. This is used to describe the rigid body transformation relationship between the radar coordinate system and the camera coordinate system.

[0072] The aforementioned initial search region represents the three-dimensional spatial region generated in the radar coordinate system after projecting the initial detection box through the preset intrinsic parameter matrix and preset extrinsic parameter matrix in the preset calibration parameters. The three-dimensional spatial coordinates within this region are represented as follows: , where t represents a certain time. The two-dimensional semantic information obtained from visual detection is transformed into a three-dimensional spatial range that can be retrieved by radar, which is used to define the target area that needs to be associated in the radar point cloud.

[0073] In this embodiment, when the initial detection box is transformed based on preset calibration parameters to obtain the initial search area, the system uses a preset intrinsic parameter matrix to back-project the pixel coordinates of the initial detection box to the three-dimensional ray direction of the camera coordinate system, and then uses a preset extrinsic parameter matrix to transform the ray region to the radar coordinate system, finally generating a three-dimensional volume region containing the possible spatial location of the target part. This region provides precise spatial constraints for radar data association, ensuring that the subsequent matching process is only carried out within a reliable range, and avoiding the misassociation of irrelevant dynamic targets as target parts.

[0074] When determining the observation state sequence based on the initial detection box, the initial search region, and multiple clusters, the system determines the clusters located within the three-dimensional region by spatially overlapping the centroids of multiple clusters with the initial search region. It then confirms that the clusters belong to the human hand by combining the semantic information of the initial detection box. Finally, it extracts the three-dimensional coordinates, radial velocity, and motion characteristics consistent with visual semantics of the clusters and generates an observation state sequence containing changes in position, velocity, and attitude.

[0075] Based on the above optional embodiments, this application embodiment achieves coordinate system alignment between the visual and radar sensing systems through preset calibration parameters, enabling the initial detection box to be transformed into a physically meaningful initial search region, thereby establishing a verifiable spatial correspondence between visual semantics and radar dynamic targets. Spatially constraining multiple clusters through the initial search region effectively eliminates interference clustering from non-target areas, improving association accuracy. Through joint judgment of the initial detection box and clusters, it ensures that the observation state sequence is generated only from radar dynamic targets that have been doubly verified and supported by visual semantics, significantly enhancing the reliability and authenticity of the observation state sequence in complex environments.

[0076] Optionally, determining the observation state sequence based on the initial detection box, the initial search region, and multiple clusters includes: determining the target cluster among the multiple clusters based on the initial search region and multiple clusters, wherein the target cluster is used to represent the dynamic point cloud cluster corresponding to the target part; and determining the observation state sequence based on the target cluster and the initial detection box.

[0077] The aforementioned target cluster represents the unique or optimal dynamic point cloud cluster that is determined to be the one corresponding to the initial detection box and located within the initial search area among multiple clusters. Its centroid and motion characteristics both satisfy the physical characteristics of the human hand, representing the true spatial representation of the target part in the radar coordinate system.

[0078] This embodiment of the application, based on an initial search area and multiple clusters, determines the target cluster by spatially overlapping the centroids of each cluster with the initial search area. Only clusters whose centroids are located within this three-dimensional region are retained. Furthermore, a preliminary screening is performed based on whether their radial velocities conform to low-speed limb movement characteristics. Finally, the unique dynamic point cloud cluster aligned with the visual semantic region is determined as the target cluster. This process achieves cross-validation between radar perception and visual semantics, eliminating background interference and non-target dynamic objects.

[0079] When determining the observation state sequence based on target clusters and initial detection boxes, the system binds the three-dimensional spatial coordinates and radial velocity information provided by the target clusters with the semantic information of the target parts carried by the initial detection boxes. By fusing three-dimensional position, velocity, and visually confirmed limb attributes, it directly generates a sequence containing the location of the target parts. Instantaneous velocity And the observation state sequence that is subsequently used as the geometric basis for attitude estimation. The generation of this sequence depends on the dual assurance of the reliability of visual semantics and the accuracy of radar ranging.

[0080] Based on the above optional embodiments, this application embodiment spatially filters multiple clusters through an initial search region, ensuring that radar dynamic targets are matched only within the three-dimensional region corresponding to visual semantics, significantly reducing the risk of false association. Determining target clusters achieves high-precision correspondence between radar point clouds and visual detection results, avoiding multi-target confusion caused by environmental interference. Through the joint information output of target clusters and initial detection boxes, an observation state sequence with three-dimensional spatial realism and visual semantic consistency is directly constructed, ensuring that gesture state representation maintains high stability and reliability even under complex lighting and background clutter environments.

[0081] Optionally, determining the target cluster among the multiple clusters based on the initial search region and multiple clusters includes: obtaining the centroid and radial velocity of the current cluster among the multiple clusters; and determining the current cluster as the target cluster in response to the fact that the centroid of the current cluster is located in the initial search region and the radial velocity satisfies a preset feature condition.

[0082] The centroid mentioned above represents the geometric center coordinates of all radar points within a cluster, used to characterize the overall position of the cluster in three-dimensional space.

[0083] The radial velocity mentioned above represents the velocity component of the target point along the radar beam direction, measured by the radar through the Doppler effect, and is used to determine whether the target has low-speed movement characteristics consistent with human limbs.

[0084] In the process of determining the target cluster among multiple clusters based on the initial search area and multiple clusters, the system first obtains the centroid and radial velocity of the current cluster in multiple clusters. That is, it extracts the three-dimensional coordinates of the centroid and the average radial velocity value of each cluster as a quantitative representation of the spatial position and motion state of the cluster, providing basic input data for subsequent matching and determination.

[0085] Then, in response to the fact that the centroid of the current cluster is located within the initial search area and the radial velocity meets the preset characteristic conditions, the current cluster is determined to be the target cluster. Specifically, the system performs the following dual determination for each current cluster: first, it determines whether its centroid falls within the three-dimensional spatial boundary of the initial search area; second, it determines whether its radial velocity is within a preset low-speed motion range, such as 0.1-1.5 m / s. The above preset characteristic conditions are set based on the typical motion speed range of the human hand in gesture interaction, excluding interference from high-speed moving vehicles or flying birds. Only when the centroid is located within the initial search area and the radial velocity meets the preset characteristic conditions is the cluster confirmed as the target cluster; otherwise, it is discarded.

[0086] Based on the above optional embodiments, this application embodiment ensures strict alignment between radar dynamic targets and visual semantic regions in three-dimensional space by determining the spatial consistency between the centroid and the initial search area. By matching radial velocity with preset feature conditions, non-human dynamic interference is eliminated, achieving semantic filtering of motion attributes. The dual constraint mechanism gives the target cluster determination process clear physical meaning and discrimination logic, avoiding mismatches caused by relying solely on a single modality. Therefore, in environments with strong light, backlight, nighttime, or cluttered backgrounds, it significantly improves the accuracy and anti-interference capability of target cluster identification, ensuring high reliability of the input source of the observation state sequence.

[0087] Optionally, the radar sensing data further includes: the three-dimensional instantaneous velocity corresponding to the target location; determining the observation state sequence based on the target cluster and the initial detection box includes: determining the target detection box based on the target cluster and the initial detection box, wherein the target detection box is used to determine the spatial coordinate position corresponding to the target location; extracting features from the target detection box to obtain the target location contour; fitting an ellipse to the target location contour to obtain a target fitting function, wherein the target fitting function is used to determine the pointing angle of the target location contour within the target detection box; and determining the observation state sequence within a preset time window based on the target detection box, the target fitting function, and the three-dimensional instantaneous velocity corresponding to the target location.

[0088] The three-dimensional instantaneous velocity corresponding to the above target location It represents the instantaneous velocity component of the target cluster as a whole in three-dimensional space calculated by radar Doppler effect. Its value comes from the vector reconstruction of the average radial velocity of all points in the target cluster in the global coordinate system, and is used to characterize the speed and direction of the target part in three-dimensional space.

[0089] The aforementioned target detection box represents the final detection area in the image plane that is determined after the target cluster and the initial detection box have completed spatial association. Its position is consistent with the initial detection box, but its semantics have been verified by radar data and it has real spatial traceability.

[0090] The aforementioned target region contour represents a sub-pixel level hand boundary curve obtained by performing skin color space segmentation (Luminance-Chroma Component Space, YCrCb) and Canny EdgeDetector detection on the image region within the target detection box. This curve is used to describe the geometric shape of the target region.

[0091] The above ellipse fitting method represents the use of the least squares method to perform nonlinear optimization on the contour C of the target part, fitting an optimal matching ellipse. The objective fitting function is defined as follows:

[0092]

[0093] in, The coordinates of the i-th pixel on the contour C of the target region are represented by the coordinates of the center of the fitted ellipse obtained through iterative optimization. The length of the major semi-axis of the fitted ellipse is a, the length of the minor semi-axis of the fitted ellipse is b, and the rotation angle θ of the fitted ellipse is θ, and θ is directly mapped to the pointing angle of the palm in the image plane.

[0094] The aforementioned preset time window represents a fixed continuous frame time interval N set by the system for constructing the observation state sequence. It can be 1.5 seconds and is used to collect and integrate multi-frame dynamic data of the target part during the execution of the gesture action to ensure that the generated observation state sequence can fully reflect the start, movement and termination stages of the gesture and avoid misjudgment due to single-frame noise or instantaneous jitter.

[0095] The above observation state sequence represents a dynamic data sequence within a preset time window, consisting of the target location's coordinates in three-dimensional space, its instantaneous three-dimensional velocity, the palm pointing angle, and its angular velocity changes. It is used to fully characterize the spatiotemporal evolution of gestures.

[0096] In this embodiment of the application, when determining the observation state sequence based on the target cluster and the initial detection box, the target detection box is first determined based on the target cluster and the initial detection box. The target cluster that has been verified by radar is semantically bound to the initial detection box to confirm that the target part corresponding to the initial detection box has real three-dimensional spatial support in the visual domain, thereby upgrading the initial detection box to a target detection box with physical credibility.

[0097] Then, feature extraction is performed on the target detection box to obtain the contour of the target area. Specifically, within the image area defined by the target detection box, the system first performs skin pixel segmentation using the YCrCb color space, then performs Canny edge detection, and finally outputs a continuous, subpixel-precision hand contour curve as input for morphological analysis.

[0098] The target region's contour is then fitted with an ellipse to obtain the target fitting function. The system uses the least squares method to perform nonlinear optimization fitting on the contour, outputting a function that includes the center point. The ellipse parameter set consisting of major axis a, minor axis b, and rotation angle θ enables the geometric quantization of gesture postures.

[0099] Finally, within a preset time window, the observation state sequence is determined based on the target detection box, the target fitting function, and the corresponding 3D instantaneous velocity of the target location. That is, the system can continuously record the target detection box position, the rotation angle θ output by the target fitting function, and the 3D instantaneous velocity provided by the target cluster in each frame, with a time window of 1.5 seconds. These parameters are combined in chronological order to form the state vector at each time step. And constitute a complete dynamic sequence of gestures within the time window. , used to describe the dynamic process of a gesture. Let be the angular velocity at time t.

[0100] Based on the above optional embodiments, this application embodiment verifies the spatial credibility of the target detection box through target clustering to ensure that the visual detection results have true three-dimensional support. By fitting the ellipse of the target part's contour, the pointing angle of the palm is extracted as a key posture feature, achieving geometric analysis of the gesture direction. By fusing the spatial position of the target detection box, the pointing angle of the target fitting function, and the three-dimensional instantaneous velocity corresponding to the target part, an observation state sequence with high dimension, high precision, and high temporal consistency is constructed within a preset time window, significantly improving the stability and semantic richness of the gesture state representation under complex lighting and background interference.

[0101] Figure 3 This is a schematic diagram of an optional feature extraction method according to an embodiment of this application, such as... Figure 3 As shown. After obtaining the 2D pixel bounding box of the hand and the dynamic target cluster, the two are compared using a pre-calibrated intrinsic parameter matrix K and extrinsic parameter matrix. Achieving cross-modal alignment, the system accurately correlates hand targets in 3D space, obtains the 3D coordinates of the target position, and maintains a positioning accuracy within ±0.05 meters. Feature extraction is then performed within the confirmed hand region (target detection box). The system uses YCrCb skin color segmentation and Canny edge detection to obtain sub-pixel-level contours, and solves for the target contour pointing angle (hand rotation angle θ) using an ellipse fitting algorithm. Simultaneously, 7-dimensional Hu invariant moments are extracted. Describe the posture and form, combined with three-dimensional velocity. With angular velocity Construct a sequence of gesture states within a 1.5-second time window.

[0102] Optionally, the preset state sequence includes a first state sequence and a second state sequence, with the parking directions corresponding to the first state sequence and the second state sequence being opposite. In response to the comparison result and the termination coordinate position satisfying preset parking conditions, parking control of the target vehicle includes: in response to the similarity distance between the observed state sequence and the first state sequence being less than or equal to a preset difference threshold, and the termination coordinate position being located within a preset safety trigger area, controlling the target vehicle to perform a parking operation based on a first direction, wherein the first direction is the parking direction corresponding to the first state sequence; or, in response to the similarity distance between the observed state sequence and the second state sequence being less than or equal to a preset difference threshold, and the termination coordinate position being located within a preset safety trigger area, controlling the target vehicle to perform a parking operation based on a second direction, wherein the second direction is the parking direction corresponding to the second state sequence.

[0103] The aforementioned first state sequence represents a reference template in the preset state sequence corresponding to the vehicle's intention to park to the left. Its characteristics are that the palm is vertical, the hand is waved rapidly along the left side of the vehicle body, the lateral speed is positive, and the pointing angle is stable at approximately 90°.

[0104] The aforementioned second state sequence represents a reference template in the preset state sequence corresponding to the vehicle's intention to park to the right. Its characteristics are: the palm is vertical, it is waved rapidly along the right side of the vehicle body, the lateral speed is negative, and the pointing angle is stable at approximately 90°.

[0105] The aforementioned similarity distance is calculated using a weighted dynamic time warping algorithm to determine the cumulative difference between the observed state sequence and the preset state sequence. Its composite distance function is designed as follows:

[0106]

[0107] in, This is a sequence of gesture states observed in real time. For pre-stored standard gesture template sequences, and Represents a three-dimensional instantaneous velocity vector sequence of observation and template sequences. and This represents the palm rotation angle sequence between the observation and the template sequence. and This represents the 7-dimensional Hu invariant moment sequence of the observation and template sequences. These are the weighting coefficients calibrated through actual vehicle testing. Represents the eigenvectors of the seven-dimensional Hu invariant moments It is used to comprehensively measure the degree of matching between speed, attitude and form.

[0108] The aforementioned preset safety trigger area refers to a three-dimensional spatial area centered on the vehicle, with the horizontal and vertical distances in front and behind within a preset range. The preset range can be within 2.5 meters in horizontal distance and 0.8 to 1.8 meters in height. Parking control is only allowed to be triggered when the gesture ends within this area to prevent unauthorized operation from a distance.

[0109] The first direction mentioned above represents the parking direction corresponding to the first state sequence, that is, the vehicle moves laterally to the left to exit the parking space. The second direction mentioned above represents the parking direction corresponding to the second state sequence, that is, the vehicle moves laterally to the right to exit the parking space.

[0110] This application embodiment responds to the comparison result and the termination coordinate position meeting preset parking conditions by performing parking control on the target vehicle. Specifically, it includes: when the system determines that the overall difference between the current gesture dynamics and the left parking exit template is within an allowable range, and the gesture termination point is located within the safe trigger area, the system confirms the user's intention to park on the left and generates a corresponding control command to drive the target vehicle to move laterally to the left to complete the parking. When the system determines that the overall difference between the current gesture dynamics and the right parking exit template is within an allowable range, and the gesture termination point is located within the safe trigger area, the system confirms the user's intention to park on the right and generates a corresponding control command to drive the target vehicle to move laterally to the right to complete the parking.

[0111] Based on the above optional embodiments, this application embodiment achieves accurate differentiation of parking intentions on the left and right sides of the vehicle through a bidirectional template design of the first state sequence and the second state sequence in the preset state sequence. A quantitative matching mechanism based on similarity distance and a preset difference threshold ensures the robustness and anti-interference capability of gesture intention recognition. Spatial verification between the termination coordinate position and the preset safety trigger area ensures that parking control is activated only when the user is within the safe interaction range. Thus, without relying on physical buttons or voice commands, it achieves highly safe and accurate bidirectional parking control based on dual constraints of gesture direction and spatial position.

[0112] Optionally, the vehicle control method further includes: determining that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the similarity distance between the observed state sequence and the first state sequence being greater than a preset difference threshold; or, determining that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the similarity distance between the observed state sequence and the second state sequence being greater than a preset difference threshold; or, determining that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the termination coordinate position not being located in a preset safety triggering area.

[0113] In this embodiment of the application, when controlling parking of a target vehicle, it is necessary to determine whether the real-time operation corresponding to the observed state sequence is an invalid operation. Specifically, invalid operations mainly include the following situations: First, in response to the similarity distance between the observed state sequence and the first state sequence being greater than a preset difference threshold, the real-time operation corresponding to the observed state sequence is determined to be an invalid operation. That is, when the combined difference between the current gesture dynamics and the left parking template exceeds the system's preset allowable range, the system determines that the gesture does not have the intention to park on the left, thereby excluding its possibility as a valid control command.

[0114] Second, in response to a similarity distance between the observed state sequence and the second state sequence exceeding a preset difference threshold, the real-time operation corresponding to the observed state sequence is determined to be an invalid operation. That is, when the combined difference between the current gesture dynamics and the right-side parking template exceeds the system's preset allowable range, the system determines that the gesture does not have the intention to park on the right, thus excluding it from the possibility of being a valid control command.

[0115] Third, in response to the termination coordinates not being located within the preset safety trigger area, the real-time operation corresponding to the observed state sequence is determined to be an invalid operation. That is, when the spatial coordinates of the target part exceed the safety zone boundary of 2.5 meters in front of and behind the vehicle and 0.8 to 1.8 meters in height when the gesture ends, the system determines that the operation occurred within the unauthorized interaction range, and regardless of whether the gesture shape matches the template, it is considered an invalid operation.

[0116] Based on the above optional embodiments, this application embodiment effectively filters out unintentional actions that deviate significantly from the preset parking template in terms of speed, posture, or shape through a comparison mechanism of similarity distance and preset difference threshold; it excludes long-distance, unauthorized, or erroneously triggered operations by spatial verification of the termination coordinate position and preset safety trigger area; and it forms a multi-dimensional and comprehensive invalid operation identification mechanism by independently triggering invalid operation judgment under three independent judgment conditions, which significantly improves the system's ability to identify interfering gestures, random actions, and unauthorized operations, and ensures that parking control is activated only under real, safe, and clearly intentional conditions.

[0117] Figure 4 This is a schematic diagram of an optional parking control according to an embodiment of this application, such as... Figure 4As shown. After acquiring the gesture state sequence within the time window, the system uses a weighted composite distance function to match pre-stored "left-side parking" and "right-side parking" templates. A valid intention is only determined when the overall similarity distance, calculated by the combined errors of speed, orientation, and shape, is below a preset threshold. Finally, the system verifies whether the gesture termination point is within a safe trigger area of ​​2.5 meters in front of and behind the vehicle and 0.8-1.8 meters in height, and confirms the parking path is unobstructed using ultrasonic data. Once all conditions are met, the system triggers the vehicle's automatic parking control logic via the CAN bus, achieving safe, precise, and accidental-touch-proof intelligent human-machine interaction. Otherwise, the system determines the gesture as invalid and remains in standby mode.

[0118] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0119] Figure 5 This is a structural block diagram of an optional vehicle control device according to an embodiment of this application, such as... Figure 5 As shown, it should be noted that this device can be used to execute the above-described vehicle control method. The device includes: an acquisition module 501, used to acquire an observed state sequence and a preset state sequence, wherein the observed state sequence represents the real-time operation change state corresponding to the target part of the target user within a preset time window, and includes the termination coordinate position corresponding to the target part; the preset state sequence represents the standard operation change state corresponding to the preset parking intention; a construction module 502, used to construct a composite distance function based on the observed state sequence and the preset state sequence, wherein the composite distance function is used to determine the similarity distance between the observed state sequence and the preset state sequence; a comparison module 503, used to compare the similarity distance and a preset difference threshold to obtain a comparison result, wherein the comparison result is used to determine the target parking intention corresponding to the observed state sequence; and a control module 504, used to perform parking control on the target vehicle in response to the comparison result and the termination coordinate position satisfying the preset parking conditions.

[0120] Optionally, the acquisition module 501 is further configured to: in response to the distance between the target user and the target vehicle being less than or equal to a preset distance threshold, acquire visual perception data and radar perception data corresponding to the target user, wherein the visual perception data and radar perception data are obtained by data collection of the target user through an on-board multi-source sensing unit; detect the visual perception data based on a target detection model to obtain an initial detection box, wherein the initial detection box is used to represent the two-dimensional bounding box corresponding to the target part, the target detection model is obtained by training on sample data and the labels corresponding to the sample data, the sample data is used to represent historical visual perception data, and the labels corresponding to the sample data are used to represent the detection boxes corresponding to the historical visual perception data; perform cluster analysis on the radar perception data to obtain multiple clusters, wherein the multiple clusters are used to determine the dynamic targets in the current environment of the target user; and determine the observation state sequence based on the initial detection box and the multiple clusters.

[0121] Optionally, the acquisition module 501 is further configured to: transform the initial detection box based on preset calibration parameters to obtain an initial search area, wherein the initial search area is used to represent the three-dimensional search area of ​​the target part in the radar coordinate system, wherein the preset calibration parameters include a preset intrinsic parameter matrix and a preset extrinsic parameter matrix; and determine the observation state sequence based on the initial detection box, the initial search area and multiple clusters.

[0122] Optionally, the acquisition module 501 is further configured to: determine the target cluster among the multiple clusters based on the initial search area and multiple clusters, wherein the target cluster is used to represent the dynamic point cloud cluster corresponding to the target part; and determine the observation state sequence based on the target cluster and the initial detection box.

[0123] Optionally, the acquisition module 501 is further configured to: acquire the centroid and radial velocity of the current cluster among multiple clusters; and determine the current cluster as the target cluster in response to the fact that the centroid of the current cluster is located in the initial search region and the radial velocity satisfies the preset feature conditions.

[0124] Optionally, the radar sensing data further includes: the three-dimensional instantaneous velocity corresponding to the target part. The acquisition module 501 is also used to: determine the target detection box based on the target cluster and the initial detection box, wherein the target detection box is used to determine the spatial coordinate position corresponding to the target part; extract features from the target detection box to obtain the target part contour; perform ellipse fitting on the target part contour to obtain the target fitting function, wherein the target fitting function is used to determine the pointing angle of the target part contour in the target detection box; and determine the observation state sequence within a preset time window based on the target detection box, the target fitting function, and the three-dimensional instantaneous velocity corresponding to the target part.

[0125] Optionally, the preset state sequence includes a first state sequence and a second state sequence, wherein the parking driving directions corresponding to the first state sequence and the second state sequence are opposite. The control module 504 is further configured to: control the target vehicle to perform a parking operation based on a first direction in response to the similarity distance between the observed state sequence and the first state sequence being less than or equal to a preset difference threshold, and the termination coordinate position being located in a preset safety trigger area; or, control the target vehicle to perform a parking operation based on a second direction in response to the similarity distance between the observed state sequence and the second state sequence being less than or equal to a preset difference threshold, and the termination coordinate position being located in a preset safety trigger area, wherein the second direction is the parking driving direction corresponding to the second state sequence.

[0126] Optionally, the vehicle control device further includes: a determination module 505, configured to: determine that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the similarity distance between the observed state sequence and the first state sequence being greater than a preset difference threshold; or, determine that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the similarity distance between the observed state sequence and the second state sequence being greater than a preset difference threshold; or, determine that the real-time operation corresponding to the observed state sequence is an invalid operation in response to the termination coordinate position not being located in a preset safety trigger area.

[0127] Embodiments of this application also provide a vehicle, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods described in various embodiments of this application when it runs.

[0128] Embodiments of this application also provide a computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0129] Embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0130] Embodiments of this application also provide a computer program product, including a non-volatile computer-readable storage medium for storing a computer program that, when executed by a processor, implements the methods in various embodiments of this application.

[0131] Embodiments of this application also provide a computer program that, when executed by a processor, implements the methods described in the various embodiments of this application.

[0132] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0133] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0134] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0135] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0136] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0137] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A vehicle control method, characterized in that, include: Acquire an observation state sequence and a preset state sequence, wherein the observation state sequence is used to represent the real-time operation change state corresponding to the target part of the target user within a preset time window, the observation state sequence includes the termination coordinate position corresponding to the target part, and the preset state sequence is used to represent the standard operation change state corresponding to the preset parking intention; Based on the observed state sequence and the preset state sequence, a composite distance function is constructed, wherein the composite distance function is used to determine the similarity distance between the observed state sequence and the preset state sequence; The similarity distance and a preset difference threshold are compared to obtain a comparison result, wherein the comparison result is used to determine the target parking intention corresponding to the observed state sequence; In response to the comparison result and the termination coordinate position satisfying the preset parking conditions, parking control is performed on the target vehicle.

2. The method according to claim 1, characterized in that, The acquisition of the observation state sequence includes: In response to the distance between the target user and the target vehicle being less than or equal to a preset distance threshold, visual perception data and radar perception data corresponding to the target user are acquired, wherein the visual perception data and the radar perception data are obtained by the vehicle-mounted multi-source sensing unit collecting data from the target user; The visual perception data is detected based on the target detection model to obtain an initial detection box, wherein the initial detection box is used to represent the two-dimensional bounding box corresponding to the target part. The target detection model is obtained by training on sample data and the labels corresponding to the sample data. The sample data is used to represent historical visual perception data, and the labels corresponding to the sample data are used to represent the detection boxes corresponding to the historical visual perception data. Cluster analysis is performed on the radar sensing data to obtain multiple clusters, wherein the multiple clusters are used to determine dynamic targets in the current environment of the target user; Based on the initial detection box and the multiple clusters, the observation state sequence is determined.

3. The method according to claim 2, characterized in that, Determining the observation state sequence based on the initial detection box and the multiple clusters includes: The initial detection box is transformed based on preset calibration parameters to obtain an initial search area, wherein the initial search area is used to represent the three-dimensional search area of ​​the target part in the radar coordinate system, and the preset calibration parameters include a preset intrinsic parameter matrix and a preset extrinsic parameter matrix. The observation state sequence is determined based on the initial detection box, the initial search region, and the multiple clusters.

4. The method according to claim 3, characterized in that, The step of determining the observation state sequence based on the initial detection box, the initial search region, and the multiple clusters includes: Based on the initial search region and the plurality of clusters, a target cluster is determined among the plurality of clusters, wherein the target cluster is used to represent the dynamic point cloud cluster corresponding to the target part; Based on the target cluster and the initial detection box, the observation state sequence is determined.

5. The method according to claim 4, characterized in that, The step of determining the target cluster among the multiple clusters based on the initial search region and the multiple clusters includes: Obtain the centroid and radial velocity of the current cluster among the multiple clusters; In response to the fact that the centroid of the current cluster is located in the initial search region and the radial velocity satisfies a preset feature condition, the current cluster is determined to be the target cluster.

6. The method according to claim 4, characterized in that, The radar sensing data further includes: the three-dimensional instantaneous velocity corresponding to the target location; determining the observation state sequence based on the target cluster and the initial detection box includes: Based on the target cluster and the initial detection box, a target detection box is determined, wherein the target detection box is used to determine the spatial coordinate position corresponding to the target part; Feature extraction is performed on the target detection box to obtain the contour of the target region; Ellipse fitting is performed on the contour of the target part to obtain a target fitting function, wherein the target fitting function is used to determine the pointing angle of the contour of the target part in the target detection box; Within the preset time window, the observation state sequence is determined based on the target detection box, the target fitting function, and the three-dimensional instantaneous velocity corresponding to the target part.

7. The method according to claim 1, characterized in that, The preset state sequence includes a first state sequence and a second state sequence, wherein the parking driving directions corresponding to the first state sequence and the second state sequence are opposite, and the parking control of the target vehicle in response to the comparison result and the termination coordinate position satisfying the preset parking conditions includes: In response to the similarity distance between the observed state sequence and the first state sequence being less than or equal to the preset difference threshold, and the termination coordinate position being located within a preset safety trigger area, the target vehicle is controlled to perform a parking operation based on a first direction, wherein the first direction is the parking driving direction corresponding to the first state sequence; or, In response to the similarity distance between the observed state sequence and the second state sequence being less than or equal to the preset difference threshold, and the termination coordinate position being located within the preset safety trigger area, the target vehicle is controlled to perform a parking operation based on the second direction, wherein the second direction is the parking driving direction corresponding to the second state sequence.

8. The method according to claim 7, characterized in that, The method further includes: In response to the similarity distance between the observed state sequence and the first state sequence being greater than the preset difference threshold, the real-time operation corresponding to the observed state sequence is determined to be an invalid operation; or, In response to the similarity distance between the observed state sequence and the second state sequence being greater than the preset difference threshold, the real-time operation corresponding to the observed state sequence is determined to be the invalid operation; or, In response to the fact that the termination coordinate position is not located in the preset safety trigger area, the real-time operation corresponding to the observation state sequence is determined to be the invalid operation.

9. A vehicle, characterized in that, include: Memory, which stores executable programs; A processor for running the program, wherein the program, when running, performs the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored executable program, wherein, when the executable program is executed, it controls the device on which the storage medium is located to perform the method according to any one of claims 1 to 8.

Images