Deep learning-based automatic parking method and device, vehicle and medium
Patent Information
- Application Number
- CN202610852657.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-12
- Publication Date
- 2026-09-25
AI Technical Summary
若车位一侧存在固定障碍物（如地下车库中的水泥立柱、墙面等刚性高高度障碍物），即使泊入过程中未发生碰撞，泊入后驾驶员侧车门与障碍物之间的横向距离也可能严重不足，导致车门最大开启角度被限制在极小范围（例如15°以下），驾驶员无法正常进出车辆
首先，本技术方案在泊入前主动检测驾驶员侧边界外的障碍物，并基于车门几何参数预判泊车终点处的最大车门开启宽度。当判断该宽度低于正常下车所需的阈值时，不依赖驾驶员的经验调整，而是自动触发优化机制。其次，方案将优化目标量化为具体的横向偏移量。根据障碍物位置及车位可用宽度，计算出一个能使车门开启宽度达标所需的横向偏移值，并据此生成一个偏向车位远离障碍物一侧的目标位姿。偏移位姿打破了常规泊车追求居中的思路，直接针对开门空间不足这一核心矛盾进行修正。最后，方案以该偏移位姿为终点重新规划路径并控制车辆执行。车辆最终停泊在非对称位置，为驾驶员侧预留出足够的开门避让空间。整个过程不增加用户操作，且通过深度学习的感知与规划能力，确保在狭窄车位中既能安全避障泊入，又能保障驾驶员正常上下车。
Smart Images
Figure CN122808706A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of automotive engineering technology, specifically to an automatic parking method, device, vehicle, and medium based on deep learning. Background Technology
[0002] Automated parking systems have become an important component of modern intelligent driving assistance functions, especially with the widespread adoption of Remote Parking Assist (RPA), which allows users to control their vehicles autonomously into narrow parking spaces via a mobile app, greatly improving parking convenience. Existing automated parking systems typically rely on sensors such as ultrasonic radar and surround-view cameras to perceive parking spaces and surrounding obstacles, and then use path planning algorithms to generate a collision-free driving trajectory, enabling the vehicle to safely park in the target location.
[0003] However, most existing parking path planning methods only focus on geometric collision avoidance between the vehicle and surrounding obstacles during movement, i.e., ensuring that the vehicle body does not contact obstacles. After the vehicle is parked at the destination, the system usually only requires the vehicle to be centered or parallel to the parking space boundary, ignoring the space requirements for occupants to open the door and get out at the final parking state. Especially for the driver's side door, the outermost point of the door will sweep across a fan-shaped envelope area during its opening process. If there is a fixed obstacle on one side of the parking space (such as a rigid, high-height obstacle such as a concrete pillar or wall in an underground garage), even if no collision occurs during parking, the lateral distance between the driver's side door and the obstacle after parking may be severely insufficient, resulting in the maximum opening angle of the door being limited to a very small range (e.g., below 15°), making it impossible for the driver to enter or exit the vehicle normally. Summary of the Invention
[0004] This invention provides an automatic parking method, device, vehicle, and medium based on deep learning, which can ensure the safe parking of a vehicle in a narrow parking space, while raising the opening angle of the driver's side door to the range required for normal exit.
[0005] This invention provides an automatic parking method based on deep learning, the method comprising: The position of the obstacle relative to the target parking space is determined by detecting an obstacle located outside the driver's side boundary of the target parking space; Based on the location of the obstacle and the geometric parameters of the current vehicle's door, the maximum door opening width of the current vehicle is determined, wherein the maximum door opening width is the maximum opening width that can be formed between the driver's side door and the obstacle after the current vehicle enters the target parking space along the current planned path; When the maximum door opening width is less than a preset width threshold, the lateral offset required to make the maximum door opening width reach the width threshold is determined based on the position of the obstacle and the available width of the target parking space. Based on the lateral offset, an offset pose is generated within the target parking space that is biased away from the obstacle. Using the offset pose as the parking endpoint, a corrected planning path is generated from the current position of the current vehicle to the target position where the offset pose is located; Control the current vehicle to travel along the corrected planned path to the target position and park at the target position in the offset pose.
[0006] Optionally, determining the position of the obstacle relative to the target parking space by detecting an obstacle located outside the driver's side boundary of the target parking space includes: The vehicle uses its onboard camera to acquire an environmental image of the environment in which the vehicle is located, and identifies target obstacles from the environmental image. The target obstacle is semantically segmented to obtain its semantic category and height. If the semantic category of the target obstacle is represented as a first obstacle and the height of the target obstacle is lower than a preset height threshold, then the target obstacle is determined to be a non-blocking obstacle, and the current vehicle is controlled to drive into the target parking space along the current planned path; If the semantic category of the target obstacle is characterized as a second obstacle or the height of the target obstacle is not lower than the preset height threshold, then the target obstacle is determined to be a blocking obstacle, and the position of the obstacle relative to the target parking space is determined.
[0007] Optionally, the specific method for obtaining the height of the target obstacle includes: The target obstacle is then subjected to three-dimensional voxelization processing based on the local map constructed from the ranging information of ultrasonic radar and the real-time visual positioning information. Extract the height distribution features of the target obstacle within the preset door swing height range from the three-dimensional voxelization processing results; The height of the target obstacle is calculated based on the height distribution characteristics.
[0008] Optionally, controlling the current vehicle to travel along the corrected planned path to the target location and parking at the target location with the offset pose includes: During the driving process, a three-dimensional environment map containing the obstacles is constructed; A preset multi-level door opening envelope model is invoked. The multi-level door opening envelope model includes multiple envelopes, and each envelope has a corresponding mapping relationship with a door opening angle. The multi-level door opening envelope model is placed in the offset pose and subjected to interference verification with the three-dimensional environment map; If the verification result indicates that any of the envelopes interferes with the obstacle under the offset pose, then the offset pose and the corrected planning path are adjusted until each of the envelopes does not interfere with the obstacle.
[0009] Optionally, the specific methods for constructing the multi-level door opening envelope model include: Based on the door opening angle, the characteristic length of the door hinge mechanism, the door width, the door height, and the hinge installation position, an opening envelope function is established, wherein the output of the opening envelope function is the swept space volume or motion envelope of the door opening process. Based on the preset driver entry and exit requirements, three opening states are set. The first opening state corresponds to the first opening angle that meets the scenario of people entering and exiting from the side. The second opening state corresponds to the second opening angle that meets the scenario of people entering and exiting normally. The third opening state corresponds to the third opening angle that meets the scenario of the car door being fully open. The second opening angle is greater than the first opening angle, and the third opening angle is greater than the second opening angle. The first opening angle, the second opening angle, and the third opening angle are respectively input into the door opening envelope function to generate a first-level envelope, a second-level envelope, and a third-level envelope in sequence. The first-level envelope, the second-level envelope, and the third-level envelope together constitute the multi-level door opening envelope model.
[0010] Optionally, determining the lateral offset required to make the maximum door opening width reach the width threshold based on the position of the obstacle and the available width of the target parking space includes: Based on the semantic category and height of the obstacle, a door opening demand weighting factor is determined on the driver's side, wherein the door opening demand weighting factor characterizes the degree of influence of the obstacle on the driver's ease of getting out of the vehicle; The weighted safe disembarkation threshold is obtained by associating the door opening demand weighting factor with the width threshold, wherein the weighted safe disembarkation threshold is greater than or equal to the width threshold. Based on the location of the obstacle and the available width of the target parking space, determine the lateral offset required to make the maximum door opening width reach the weighted safe exit threshold.
[0011] Optionally, during the process of controlling the current vehicle to travel along the corrected planned path, the deep learning-based automatic parking method further includes: When the current vehicle is in a remote parking scenario, obtain the location of the person controlling the vehicle outside the vehicle relative to the target parking space; Based on the location of the person, the location of the obstacle, and the real-time position of the current vehicle, determine whether the current vehicle traveling along the corrected planned path will cause the vehicle body or an open door to squeeze the external controller onto the obstacle. If so, control the current vehicle to perform a safety braking operation.
[0012] The present invention also provides an automatic parking device based on deep learning, the device comprising: The detection module is used to determine the position of an obstacle relative to the target parking space by detecting obstacles located outside the driver's side boundary of the target parking space; The calculation module is used to determine the maximum door opening width of the current vehicle based on the position of the obstacle and the door geometry parameters of the current vehicle. The maximum door opening width is the maximum opening width that can be formed between the driver's side door and the obstacle after the current vehicle enters the target parking space along the current planned path. An estimation module is used to determine the lateral offset required to make the maximum door opening width reach the width threshold based on the position of the obstacle and the available width of the target parking space when the maximum door opening width is less than a preset width threshold. The generation module is used to generate an offset pose within the target parking space that is biased away from the obstacle, based on the lateral offset. The planning module is used to generate a corrected planning path from the current position of the current vehicle to the target position where the offset pose is located, with the offset pose as the parking endpoint. The control module is used to control the current vehicle to travel along the corrected planned path to the target position and park at the target position in the offset pose.
[0013] The present invention also provides a vehicle, the vehicle including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the deep learning-based automatic parking method as described in any of the preceding claims.
[0014] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the deep learning-based automatic parking method as described in any of the preceding claims.
[0015] The present invention has at least the following beneficial effects: First, this technical solution proactively detects obstacles outside the driver's side boundary before parking and predicts the maximum door opening width at the parking endpoint based on the door's geometric parameters. When this width is determined to be lower than the threshold required for normal exiting the vehicle, an optimization mechanism is automatically triggered without relying on the driver's experience for adjustment. Second, the solution quantifies the optimization target into a specific lateral offset. Based on the obstacle's position and the available width of the parking space, a lateral offset value required to achieve the required door opening width is calculated, and a target pose biased towards the side of the parking space away from the obstacle is generated accordingly. This offset pose breaks away from the conventional parking approach of aiming for centering, directly addressing the core contradiction of insufficient door opening space. Finally, the solution replans the path and controls the vehicle's execution using this offset pose as the endpoint. The vehicle ultimately parks in an asymmetrical position, reserving sufficient space for the driver to open the door and avoid obstacles. The entire process does not increase user operation and, through deep learning's perception and planning capabilities, ensures safe obstacle avoidance parking in narrow parking spaces while also guaranteeing normal driver entry and exit. Attached Figure Description
[0016] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.
[0017] Figure 1 This is a flowchart illustrating the steps of an automated parking method based on deep learning. Figure 2 This is a flowchart of step S101 in an automatic parking method based on deep learning; Figure 3 This is a flowchart of step S103 in a deep learning-based automatic parking method. Figure 4 This is a flowchart of step S106 in a deep learning-based automatic parking method. Figure 5 This is a schematic diagram of door opening envelope geometry modeling in a deep learning-based automatic parking method; Figure 6 This is another schematic diagram of door opening envelope geometry modeling in a deep learning-based automatic parking method; Figure 7 This is a schematic diagram illustrating a scenario for identifying the risk of people being squeezed in an automated parking method based on deep learning. Figure 8 This is a schematic diagram of the architecture of an automated parking system based on deep learning; Figure 9 This is a schematic diagram of an automatic parking device based on deep learning. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0019] Researchers in this application found that current automatic parking assistance and remote parking systems primarily rely on ultrasonic radar and surround-view cameras for obstacle detection. Their core objective is usually limited to achieving collision-free parking, i.e., planning an obstacle avoidance path by sensing the parking space boundary lines and surrounding obstacles. However, in practical applications, especially in the frequently occurring narrow parking space scenarios, existing technologies have significant functional limitations. Most existing algorithms only focus on the static distance between the vehicle's external outline and obstacles, without considering the physical space required for the vehicle doors to open after parking. This results in the system being able to perfectly avoid obstacles and park, but often causing the door opening angle to be too small or even impossible to open, preventing the driver from getting out of the vehicle or entering the car, thus rendering the remote parking function useless. At the same time, ultrasonic radar may enter detection blind spots or generate residual vibration signals at close range, making it difficult to accurately distinguish the height, material, and geometry of obstacles. The system cannot determine the actual impact of low curbs and tall walls on door opening. In addition, existing path planning algorithms often aim for the parking space to be centered, which can restrict the driver's side's movement space when the attributes of obstacles on both sides are asymmetrical. Drastic changes in lighting or environmental disturbances such as reflections from water on the ground can easily cause the system to misjudge shadows as obstacles or fail to recognize flexible features such as bushes, triggering unnecessary emergency braking and seriously affecting the user experience.
[0020] To address the aforementioned technical problems, this technical solution provides a deep learning-based automatic parking method, device, vehicle, and medium that can ensure safe parking of a vehicle in narrow spaces while simultaneously raising the driver's side door opening angle to the range required for normal exit. The following are various embodiments of this technical solution.
[0021] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of an automated parking method based on deep learning.
[0022] This embodiment provides an automatic parking method based on deep learning, including: S101. By detecting obstacles located outside the driver's side boundary of the target parking space, determine the position of the obstacle relative to the target parking space.
[0023] S102. Based on the location of the obstacle and the geometric parameters of the current vehicle's door, determine the maximum door opening width of the current vehicle. The maximum door opening width is the maximum opening width that can be formed between the driver's side door and the obstacle after the current vehicle enters the target parking space along the current planned path.
[0024] S103. When the maximum door opening width is less than the preset width threshold, determine the lateral offset required to make the maximum door opening width reach the width threshold based on the position of the obstacle and the available width of the target parking space.
[0025] S104. Generate an offset pose within the target parking space that is biased away from the obstacle based on the lateral offset.
[0026] S105. Using the offset pose as the parking endpoint, generate a corrected planned path from the current position of the current vehicle to the target position where the offset pose is located.
[0027] S106. Control the current vehicle to travel along the corrected planned path to the target position and park at the target position with an offset posture.
[0028] Understandably, firstly, this technical solution proactively detects obstacles outside the driver's side boundary before parking and predicts the maximum door opening width at the parking endpoint based on the door's geometric parameters. When this width is determined to be lower than the threshold required for normal exiting the vehicle, the optimization mechanism is automatically triggered without relying on the driver's experience for adjustment. Secondly, the solution quantifies the optimization target into a specific lateral offset. Based on the obstacle's position and the available width of the parking space, a lateral offset value required to achieve the required door opening width is calculated, and a target pose biased towards the side of the parking space away from the obstacle is generated accordingly. This offset pose breaks away from the conventional parking approach of pursuing centering, directly addressing the core contradiction of insufficient door opening space. Finally, the solution replans the path and controls the vehicle execution using this offset pose as the endpoint. The vehicle ultimately parks in an asymmetrical position, reserving sufficient space for the driver to open the door and avoid obstacles. The entire process does not increase user operation and, through the perception and planning capabilities of deep learning, ensures safe obstacle avoidance parking in narrow parking spaces while also guaranteeing normal driver entry and exit.
[0029] Please refer to Figure 2 , Figure 2 This is a flowchart of step S101 in an automatic parking method based on deep learning.
[0030] In some embodiments, step S101 includes: S201. Use the vehicle's onboard camera to acquire an environmental image of the current vehicle's environment and identify target obstacles from the environmental image.
[0031] S202. Perform semantic segmentation on the target obstacle to obtain its semantic category and height.
[0032] S203. If the semantic category of the target obstacle is represented as the first obstacle and the height of the target obstacle is lower than the preset height threshold, then the target obstacle is determined to be a non-blocking obstacle, and the current vehicle is controlled to drive into the target parking space along the current planned path.
[0033] S204. If the semantic category of the target obstacle is represented as a second obstacle or the height of the target obstacle is not lower than a preset height threshold, then the target obstacle is determined to be a blocking obstacle, and the position of the obstacle relative to the target parking space is determined.
[0034] Understandably, this solution can intelligently distinguish between obstructive and non-obstructive obstacles. The offset parking strategy is only triggered when there is a sufficiently high and rigid obstructive obstacle, such as a pillar or wall, next to the parking space; for low or traversable obstacles, the normal parking path is maintained. This function avoids unnecessary path adjustments, improves the system's adaptability to different scenarios and the rationality of its decisions, and reduces meaningless calculations and actions while ensuring the driver's side door can be opened normally, making the automatic parking process more efficient, accurate, and user-friendly.
[0035] In some embodiments, the specific methods for obtaining the height of the target obstacle include: A local map is constructed based on the ranging information of ultrasonic radar and the real-time visual positioning information. The target obstacle is then processed into a three-dimensional voxel. The height distribution characteristics of the target obstacle within the preset door swing height range are extracted from the three-dimensional voxel processing results. The height of the target obstacle is then calculated based on the height distribution characteristics.
[0036] Understandably, this solution can construct a local map based on the fusion of ultrasonic radar and visual information, perform 3D voxelization processing on obstacles, and extract height distribution features within the door swing height range, thereby accurately calculating the height of the target obstacle. This approach avoids potential misjudgments that may occur with visual recognition alone, enabling the system to accurately distinguish between high-rigidity obstructing obstacles and low-lying non-obstructing obstacles. This provides a reliable basis for triggering the offset parking strategy, further improving the decision-making accuracy and safety of automatic parking in complex scenarios.
[0037] Please refer to Figure 3 , Figure 3 This is a flowchart of step S103 in an automatic parking method based on deep learning.
[0038] In some embodiments, step S103 includes: S301. Based on the semantic category and height of the obstacle, determine the door opening demand weight factor on the driver's side, wherein the door opening demand weight factor characterizes the degree of influence of the obstacle on the driver's convenience of getting out of the vehicle.
[0039] S302. Perform a correlation calculation between the door opening demand weight factor and the width threshold to obtain the weighted safe disembarkation threshold, wherein the weighted safe disembarkation threshold is greater than or equal to the width threshold.
[0040] S303. Based on the location of the obstacle and the available width of the target parking space, determine the lateral offset required to make the maximum door opening width reach the weighted safe exit threshold.
[0041] Understandably, this solution can dynamically adjust the safe exit threshold based on the semantic category and height of the obstacle. For obstacles with high impact, such as pillars and walls, the system automatically increases the threshold requirement and calculates a larger lateral offset to provide the driver with more space to exit; for obstacles with less impact, the standard is used. This adaptive mechanism makes parking posture optimization more targeted and flexible, ensuring the driver can exit normally while avoiding excessive adjustments that occupy adjacent parking space, achieving an intelligent balance between exit convenience and space utilization.
[0042] In some embodiments, step S104 includes: The lateral displacement of the current vehicle relative to the default centering pose is determined based on the lateral offset. It is then determined whether the current vehicle exceeds the boundary of the target parking space while maintaining the default parking posture after applying the lateral displacement. If it exceeds the boundary of the target parking space, a yaw angle adjustment is determined so that when the current vehicle is parked in an offset pose composed of the lateral displacement and the yaw angle adjustment, the overall outline of the current vehicle remains within the boundary range of the target parking space, and the opening gap between the driver's side door and the obstacle meets the threshold requirement.
[0043] Understandably, this solution considers not only lateral displacement but also yaw angle adjustment when generating the offset pose. The system determines whether simple lateral displacement would cause the vehicle to cross the boundary. If it does, it fine-tunes the vehicle's attitude angle to ensure the vehicle is safely parked within the parking space boundary, while simultaneously ensuring the driver's side door opening clearance meets the standard. This mechanism resolves the contradiction between insufficient offset and the risk of crossing the boundary in narrow parking spaces. It maximizes space utilization through asymmetric tilt parking within a limited space, ensuring that the vehicle can be parked comfortably and the door can be fully opened.
[0044] In some embodiments, step S105 includes: Obtain the boundary range information of the target parking space, which includes at least the corner coordinates of the target parking space; establish the vehicle outline model based on the current vehicle's external dimensions; during the generation of the corrected planning path, the path nodes are collision detected and filtered under the constraint that the pose of the vehicle outline model at each moment of the entire path does not exceed the boundary range of the target parking space, until a corrected planning path that meets the constraint is generated.
[0045] Understandably, this solution uses the rigid constraint that the vehicle's outline does not exceed the parking space boundary throughout the entire planned correction path. Based on the parking space corner coordinates and the vehicle's size model, the system performs collision detection and screening on the vehicle's pose at various moments along the path, ensuring that the vehicle does not interfere with obstacles or adjacent vehicles outside the parking space lines as it moves towards the offset endpoint. This mechanism guarantees the safety and feasibility of the asymmetric parking path, avoiding the introduction of new collision risks due to adjustments in the final pose, making the tilting parking process in narrow parking spaces both accurate and safe.
[0046] Please refer to Figure 4 , Figure 4 This is a flowchart of step S106 in a deep learning-based automatic parking method.
[0047] In some embodiments, step S106 includes: S401. During driving, construct a three-dimensional environment map containing obstacles.
[0048] S402. Call the preset multi-level door opening envelope model. The multi-level door opening envelope model includes multiple envelopes, and each envelope has a corresponding mapping relationship with a door opening angle.
[0049] S403. Place the multi-level door opening envelope model in an offset pose and perform interference verification with the 3D environment map.
[0050] S404. If the verification result shows that any envelope interferes with an obstacle in the offset pose, then adjust the offset pose and the corrected planning path until each envelope does not interfere with an obstacle.
[0051] Understandably, this solution utilizes a multi-level door opening envelope model and a 3D environment map for interference verification before parking and during driving. The system verifies the presence of interference risks under different door opening angles using multiple envelopes corresponding to different door opening angles. If interference occurs, the system dynamically adjusts the pose and path until all door opening angles are safe and collision-free. This mechanism verifies and ensures the driver's side door's full-angle opening capability in advance at a 3D spatial level, completely eliminating the potential problem of the door failing to open properly after parking, thus improving the safety, reliability, and user experience of the automatic parking system.
[0052] In some embodiments, the specific methods for performing interference verification include: In the multi-level door opening envelope model, each envelope is spatially interfered with the 3D environment map to obtain the minimum distance between the envelope and the obstacle. If the minimum distance corresponding to the target envelope is less than the preset safe buffer distance value when the target envelope is in the open state, it is determined that the target envelope interferes with the obstacle.
[0053] In some embodiments, the specific methods for constructing a multi-level door opening envelope model include: Based on the door opening angle, the characteristic length of the door hinge mechanism, the door width, the door height, and the hinge installation position, an opening envelope function is established. The output of the opening envelope function is the swept space volume or motion envelope during the door opening process. Based on the preset driver entry and exit requirements, three levels of opening states are set. The first level of opening state corresponds to the first opening angle that satisfies the scenario of people entering and exiting from the side. The second level of opening state corresponds to the second opening angle that satisfies the scenario of people entering and exiting normally. The third level of opening state corresponds to the third opening angle that satisfies the scenario of the door being fully open. The second opening angle is greater than the first opening angle, and the third opening angle is greater than the second opening angle. The first, second, and third opening angles are respectively input into the opening envelope function to generate the first-level envelope, the second-level envelope, and the third-level envelope in sequence. The first-level envelope, the second-level envelope, and the third-level envelope together constitute a multi-level opening envelope model.
[0054] Geometric modeling of multi-level door opening envelope model, such as Figure 5 , 6 As shown in the figure, 201 represents the main body of the vehicle body (left front door area); 202 represents the center point of the door hinge (origin P); 203 represents the door length (radius L); 204 represents the dynamic door opening envelope (sweeped area); 205 represents the side profile of the door; 206 represents the reference ground plane; 207 represents ground obstacles (used to verify the validity of the ground clearance H). W represents the door panel thickness; H represents the ground clearance; P represents the center point of the door hinge; θ represents the door opening angle; and L represents the door length.
[0055] Please refer to Figure 7 , Figure 7 This is a schematic diagram illustrating a scenario for identifying the risk of people being squeezed in an automated parking method based on deep learning.
[0056] In some embodiments, during the process of controlling the current vehicle to travel along the corrected planned path, a deep learning-based automatic parking method further includes: When the vehicle is in a remote parking scenario, the location of the external controller relative to the target parking space is obtained. Based on the location of the person, the location of the obstacle, and the real-time pose of the current vehicle, it is determined whether the current vehicle traveling along the corrected planned path will cause the vehicle body or the open door to squeeze the external controller onto the obstacle. If so, the current vehicle is controlled to perform a safety braking operation.
[0057] Understandably, this solution introduces a safety protection mechanism for the operator outside the vehicle in remote parking scenarios. The system proactively predicts the risk of the operator being trapped between the vehicle and obstacles by real-time location tracking, combined with the vehicle's position and obstacle distribution, and executes safety braking before the risk occurs. This mechanism effectively addresses the safety hazard of users potentially entering dangerous areas when remotely parking from outside the vehicle, and is particularly suitable for narrow parking spaces. It provides proactive anti-pinch protection for operators outside the vehicle, significantly improving the safety of remote automatic parking.
[0058] This technical solution also provides a specific embodiment for implementing the above-mentioned deep learning-based automatic parking method.
[0059] In a typical scenario of a narrow parking space in an underground garage, there is a concrete pillar on the left and a large sedan on the right. When the user activates the remote parking function from outside the car, the vehicle begins to slowly park. The surround-view camera identifies a rigid, tall concrete pillar on the left and a rigid, medium-height adjacent vehicle on the right. Based on the vehicle's parameters, the algorithm calculates that the maximum opening angle of the left front door after parking along the current path is only 15 degrees, making it impossible to exit the car due to the pillar's obstruction. The planning module recognizes that the adjacent car on the right has a shorter front end and there is extra space to the right of the parking space. Therefore, it automatically corrects the path, causing the vehicle to tilt slightly to the right when parking in the final position, achieving an asymmetrical tail swing, thereby increasing the relative distance between the left front door and the pillar. After optimization, the opening angle of the left front door is increased to 35 degrees. At the same time, the mobile application sends a notification to the user that the parking position has been optimized, and the driver can exit the car normally from the driver's side. In addition, during the parking process, if the user accidentally walks into the gap between the vehicle and the pillar due to a blind spot, the system will detect the risk of the person's position and immediately stop the vehicle, while simultaneously sounding an alarm.
[0060] Understandably, this technical solution primarily addresses the core pain point of remote parking in narrow parking spaces: the car can park but the driver cannot exit. By introducing deep learning-based semantic recognition and dynamic door opening envelope modeling, the system gains the ability to perceive spatial availability. The resulting benefits include: improved safety and usability of the expected function; the system anticipates door opening interference risks and provides a score before parking; if it predicts the inability to exit, it proactively warns and guides the user to change parking spaces, avoiding potential risks of people being trapped or squeezed in narrow gaps; asymmetric path optimization; unlike traditional centering parking, the system automatically calculates lateral offset based on the attributes of obstacles on both sides, reserving maximum door opening space on the driver's side while ensuring no collision; improved multi-dimensional perception accuracy; using deep learning for semantic segmentation to accurately distinguish obstacle height and material, dynamically adjusting door opening envelope confidence and reducing false triggers; and improved human-machine interaction loop; real-time display of predicted opening angles for all four doors via a mobile application, enhancing user trust in the remote parking system and resolving information asymmetry issues when operating from outside the vehicle.
[0061] Please refer to Figure 8 , Figure 8 which is a schematic structural diagram of an automatic parking system based on deep learning.
[0062] In this embodiment, a cross-domain collaborative system architecture includes a perception layer, a decision modeling layer, a path planning layer and an interaction layer.
[0063] At the perception level, the present solution implements multi-source perception fusion based on deep learning. The system acquires environmental images through a surround-view camera, and uses a lightweight convolutional neural network such as an improved MobileNet-V3 to perform real-time semantic segmentation. The identified categories include adjacent vehicles, wall bodies, cylinders, square columns, fire hydrants, low curbs, flexible shrubs and the like. On this basis, the system fuses the ranging information of ultrasonic radars with the local map generated by visual simultaneous localization and mapping technology, performs three-dimensional voxelization processing on obstacles on both sides of the parking space, and focuses on extracting the height distribution characteristics of the obstacles in the door swing height interval, that is, H∈ [0.3m,1.5m].
[0064] In the decision modeling layer, the present solution establishes a dynamic modeling logic for door opening envelope, namely DOE. First, the door envelope function is defined as V(door)=f(θ,L,W,H,hinge_pos), where θ is the core variable of door opening angle, that is, hinge rotation angle, L is the characteristic length of the hinge mechanism, W is the door width or lateral installation spacing, H is the door height, and hinge_pos is the hinge installation position. Three-level opening states are set based on this function: Level 1 (Access Range) corresponds to θ≈30°, which meets the minimum width required for personnel to enter and exit sideways; Level 2 (Comfort Range) corresponds to θ≈60°, which meets the requirement for normal personnel entry and exit; Level 3 (Full Access) corresponds to θ≈90°. Finally, through an interference check algorithm, real-time Boolean operation or distance field calculation is performed between V(door) and the 3D voxel map output by the perception layer, and the judgment rule is: if Dist(V(door)(θ),Obstacle)<Safety_Buffer, that is, when the minimum distance between the movement contour of the door and the obstacle at any opening angle θ is less than the specified safety buffer gap, it is determined that interference occurs.
[0065] At the planning level, the present solution adopts an asymmetric path planning optimization strategy. First, the traditional parking cost function is reconstructed, the original function is , on the basis of which a door opening envelope weight term is introduced to obtain the optimized cost function . When the system recognizes that the left side is a rigid wall and the space is tight, it will automatically calculate an offset vector For example, if there is enough space on the right, the vehicle will be shifted to the right by 10 to 15 centimeters to park, so that the driver's side door on the left can be opened to Level 1, which is about 30 degrees or more.
[0066] At the control level, this solution implements risk degrading and personnel protection mechanisms. The system utilizes ultra-wideband positioning technology or pedestrian detection algorithms to monitor the driver's coordinates relative to the parking space in real time during remote parking scenarios. If, based on the predicted parking trajectory, the system determines that the vehicle or an open door will trap personnel against lateral obstacles, the system will immediately execute a safety braking operation.
[0067] Please refer to Figure 9 , Figure 9 This is a schematic diagram of an automatic parking device based on deep learning.
[0068] This embodiment also provides an automatic parking device based on deep learning, including: The detection module 301 is used to determine the position of an obstacle relative to the target parking space by detecting obstacles located outside the driver's side boundary of the target parking space.
[0069] The calculation module 302 is used to determine the maximum door opening width of the current vehicle based on the position of the obstacle and the geometric parameters of the current vehicle door. The maximum door opening width is the maximum opening width that can be formed between the driver's side door and the obstacle after the current vehicle enters the target parking space along the current planned path.
[0070] The estimation module 303 is used to determine the lateral offset required to make the maximum door opening width reach the width threshold based on the position of the obstacle and the available width of the target parking space when the maximum door opening width is less than the preset width threshold.
[0071] The generation module 304 is used to generate an offset pose within the target parking space that is biased away from the obstacle based on the lateral offset.
[0072] The planning module 305 is used to generate a corrected planned path from the current position of the current vehicle to the target position where the offset pose is located, with the offset pose as the parking endpoint.
[0073] The control module 306 is used to control the current vehicle to travel along the corrected planned path to the target position and park at the target position with an offset posture.
[0074] This invention also provides a vehicle control device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the deep learning-based automatic parking method of the above embodiments.
[0075] Taking the example of a processor and memory in a vehicle controller being connected via a bus, the memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the control processor, and these remote memories can be connected to the control device via a network. The non-transitory software programs and instructions required to implement the control methods of the above embodiments are stored in the memory, and when executed by the processor, the control methods of the above embodiments are performed.
[0076] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.
[0077] This invention also provides a vehicle, including the vehicle control device described in the above embodiments.
[0078] The vehicle can be a private car, such as a sedan, SUV, MPV, or pickup truck. It can also be a commercial vehicle, such as a van, bus, small truck, or large semi-trailer. The vehicle must have an electric motor capable of outputting power or acting as a generator to store mechanical energy. When the vehicle is a new energy vehicle, it can be a hybrid or a pure electric vehicle.
[0079] Since the vehicle applies all the technical solutions of the above-mentioned control device or vehicle controller, it has at least all the beneficial effects brought about by the technical solutions of the above embodiments, which will not be repeated here.
[0080] Furthermore, one embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions for performing the aforementioned deep learning-based automatic parking method.
[0081] It is worth noting that, since the computer-readable storage medium of the present invention is capable of executing the deep learning-based automatic parking method of any of the above embodiments, the specific implementation and technical effects of the computer-readable storage medium of the present invention can be referred to the specific implementation and technical effects of the deep learning-based automatic parking method of any of the above embodiments.
[0082] Furthermore, one embodiment of the present invention also provides a computer program product, including a computer program or computer instructions, which are stored in a computer-readable storage medium. A processor of a computer device reads the computer program or computer instructions from the computer-readable storage medium and executes the computer program or computer instructions, causing the computer device to perform the aforementioned deep learning-based automatic parking method.
[0083] It is worth noting that, since the computer program product of this embodiment can execute the deep learning-based automatic parking method of any of the above embodiments, the specific implementation method and technical effect of the computer program product of this embodiment can refer to the specific implementation method and technical effect of the deep learning-based automatic parking method of any of the above embodiments.
[0084] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically include computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.
[0085] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.
Claims
1. A deep learning-based automatic parking method, characterized in that, The method includes: The position of the obstacle relative to the target parking space is determined by detecting an obstacle located outside the driver's side boundary of the target parking space; Based on the location of the obstacle and the geometric parameters of the current vehicle's door, the maximum door opening width of the current vehicle is determined, wherein the maximum door opening width is the maximum opening width that can be formed between the driver's side door and the obstacle after the current vehicle enters the target parking space along the current planned path; When the maximum door opening width is less than a preset width threshold, the lateral offset required to make the maximum door opening width reach the width threshold is determined based on the position of the obstacle and the available width of the target parking space. Based on the lateral offset, an offset pose is generated within the target parking space that is biased away from the obstacle. Using the offset pose as the parking endpoint, a corrected planning path is generated from the current position of the current vehicle to the target position where the offset pose is located; Control the current vehicle to travel along the corrected planned path to the target position and park at the target position in the offset pose.
2. The method according to claim 1, characterized in that, The step of determining the position of an obstacle relative to the target parking space by detecting an obstacle located outside the driver's side boundary of the target parking space includes: The vehicle uses its onboard camera to acquire an environmental image of the environment in which the vehicle is located, and identifies target obstacles from the environmental image. The target obstacle is semantically segmented to obtain its semantic category and height. If the semantic category of the target obstacle is represented as a first obstacle and the height of the target obstacle is lower than a preset height threshold, then the target obstacle is determined to be a non-blocking obstacle, and the current vehicle is controlled to drive into the target parking space along the current planned path; If the semantic category of the target obstacle is characterized as a second obstacle or the height of the target obstacle is not lower than the preset height threshold, then the target obstacle is determined to be a blocking obstacle, and the position of the obstacle relative to the target parking space is determined.
3. The method according to claim 2, characterized in that, The specific methods for obtaining the height of the target obstacle include: The target obstacle is then subjected to three-dimensional voxelization processing based on the local map constructed from the ranging information of ultrasonic radar and the real-time visual positioning information. Extract the height distribution features of the target obstacle within the preset door swing height range from the three-dimensional voxelization processing results; The height of the target obstacle is calculated based on the height distribution characteristics.
4. The method according to claim 1, characterized in that, The control of the current vehicle to travel along the corrected planned path to the target position and to park at the target position in the offset pose includes: During the driving process, a three-dimensional environment map containing the obstacles is constructed; A preset multi-level door opening envelope model is invoked. The multi-level door opening envelope model includes multiple envelopes, and each envelope has a corresponding mapping relationship with a door opening angle. The multi-level door opening envelope model is placed in the offset pose and subjected to interference verification with the three-dimensional environment map; If the verification result indicates that any of the envelopes interferes with the obstacle under the offset pose, then the offset pose and the corrected planning path are adjusted until each of the envelopes does not interfere with the obstacle.
5. The method according to claim 4, characterized in that, The specific methods for constructing the multi-level door opening envelope model include: Based on the door opening angle, the characteristic length of the door hinge mechanism, the door width, the door height, and the hinge installation position, an opening envelope function is established, wherein the output of the opening envelope function is the swept space volume or motion envelope of the door opening process. Based on the preset driver entry and exit requirements, three opening states are set. The first opening state corresponds to the first opening angle that meets the scenario of people entering and exiting from the side. The second opening state corresponds to the second opening angle that meets the scenario of people entering and exiting normally. The third opening state corresponds to the third opening angle that meets the scenario of the car door being fully open. The second opening angle is greater than the first opening angle, and the third opening angle is greater than the second opening angle. The first opening angle, the second opening angle, and the third opening angle are respectively input into the door opening envelope function to generate a first-level envelope, a second-level envelope, and a third-level envelope in sequence. The first-level envelope, the second-level envelope, and the third-level envelope together constitute the multi-level door opening envelope model.
6. The method according to claim 2 or 3, characterized in that, The step of determining the lateral offset required to make the maximum door opening width reach the width threshold based on the position of the obstacle and the available width of the target parking space includes: Based on the semantic category and height of the obstacle, a door opening demand weighting factor is determined on the driver's side, wherein the door opening demand weighting factor characterizes the degree of influence of the obstacle on the driver's ease of getting out of the vehicle; The weighted safe disembarkation threshold is obtained by associating the door opening demand weighting factor with the width threshold, wherein the weighted safe disembarkation threshold is greater than or equal to the width threshold. Based on the location of the obstacle and the available width of the target parking space, determine the lateral offset required to make the maximum door opening width reach the weighted safe exit threshold.
7. The method according to claim 1, characterized in that, In controlling the current vehicle to travel along the revised planned path, the method further includes: When the current vehicle is in a remote parking scenario, obtain the location of the person controlling the vehicle outside the vehicle relative to the target parking space; Based on the location of the person, the location of the obstacle, and the real-time position of the current vehicle, determine whether the current vehicle traveling along the corrected planned path will cause the vehicle body or an open door to squeeze the external controller onto the obstacle. If so, control the current vehicle to perform a safety braking operation.
8. An automatic parking device based on deep learning, characterized in that, The device includes: The detection module is used to determine the position of an obstacle relative to the target parking space by detecting obstacles located outside the driver's side boundary of the target parking space; The calculation module is used to determine the maximum door opening width of the current vehicle based on the position of the obstacle and the door geometry parameters of the current vehicle. The maximum door opening width is the maximum opening width that can be formed between the driver's side door and the obstacle after the current vehicle enters the target parking space along the current planned path. An estimation module is used to determine the lateral offset required to make the maximum door opening width reach the width threshold based on the position of the obstacle and the available width of the target parking space when the maximum door opening width is less than a preset width threshold. The generation module is used to generate an offset pose within the target parking space that is biased away from the obstacle, based on the lateral offset. The planning module is used to generate a corrected planning path from the current position of the current vehicle to the target position where the offset pose is located, with the offset pose as the parking endpoint. The control module is used to control the current vehicle to travel along the corrected planned path to the target position and park at the target position in the offset pose.
9. A vehicle, characterized in that, The vehicle includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the deep learning-based automatic parking method according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based automatic parking method as described in any one of claims 1 to 7.