A vehicle path planning method and device based on laser radar and camera

Abstract

This application provides a vehicle path planning method and apparatus based on LiDAR and camera. The method includes: performing obstacle fusion perception and semantic classification on perception data collected by LiDAR and camera equipment to determine the danger level of obstacles; during the node expansion process of the improved A* algorithm, determining the dynamic heuristic cost based on the danger level and the distance between nodes to be processed, determining the improved actual cost based on the path length and turning angle, and determining the potential field constraint cost of the path inflection point based on the repulsive force exerted by obstacles within the visible range on the path inflection point; traversing the nodes to be processed based on the improved actual cost, the dynamic heuristic cost, and the potential field constraint cost of the path inflection point to determine the parent node and the target node; and traversing back from the target node to the parent node to determine the target driving path of the target vehicle. This method improves the reliability and adaptability of path planning.

Description

Technical Field

[0001] This application relates to the field of vehicle path planning technology, and in particular to a vehicle path planning method and apparatus based on lidar and camera. Background Technology

[0002] Path planning is one of the core technologies for autonomous vehicles. Its core objective is to plan a safe, feasible, smooth, and efficient driving path based on the vehicle's starting and ending points in an environment with obstacles, providing a reliable basis for vehicle driving control. Currently, the A* algorithm, as an optimal path search algorithm with a heuristic strategy, is widely used in the field of vehicle path planning. It combines the global optimality guarantee of Dijkstra's algorithm with the efficient search characteristics of best-priority search, and achieves path search in a grid map by filtering path nodes through total value.

[0003] Traditional A* algorithms still have many shortcomings in practical vehicle applications: They suffer from limited perception information and insufficient safety, relying solely on grid maps for path planning and treating dynamic high-risk obstacles like pedestrians and vehicles indiscriminately with static low-risk obstacles like walls. This easily leads to planned paths approaching high-risk obstacles, posing collision risks. Furthermore, they exhibit poor path smoothness and insufficient adaptability, using only path length as the core cost indicator during node expansion. This results in numerous sharp turns and redundant inflection points, failing to conform to vehicle dynamics and leading to poor vehicle stability and low path tracking accuracy. They also suffer from weak dynamic adaptability and insufficient foresight, employing a static planning model that focuses only on obstacles within the neighborhood, without constraints on visible obstacles outside the neighborhood. This results in path inflection points easily approaching obstacles, increasing the risk of collisions. Finally, their fixed cost function makes it difficult to balance efficiency and accuracy. The heuristic cost uses fixed weights, resulting in insufficient search accuracy in complex high-risk environments and low search efficiency in open, low-risk environments, making them unsuitable for all-weather, complex, and dynamic vehicle driving scenarios. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a vehicle path planning method and device based on LiDAR and camera. The method obtains the semantic category and spatial location of obstacles through the fusion perception of LiDAR and camera, determines the danger level of obstacles based on the semantic category, and introduces the dynamic heuristic cost corresponding to the danger level, the additional cost of turning angle, and the cost of potential field constraint into the cost function of the improved A* algorithm. Then, the improved A* algorithm is used to determine the target driving path of the vehicle. This method effectively solves the technical defects of the traditional A* algorithm, such as single perception information, poor path smoothness, weak adaptability to dynamic environment, and the inability to balance search efficiency and planning accuracy. It realizes safe, efficient and smooth path planning for vehicles in complex dynamic environments, significantly improves the reliability and adaptability of path planning, and fits various practical application scenarios of autonomous vehicles.

[0005] This application provides a vehicle path planning method based on LiDAR and a camera, the method comprising: The LiDAR and camera equipment installed on the target vehicle are used to collect perception data under a unified spatiotemporal reference corresponding to the environment in which the target vehicle is located. The perception data is then subjected to obstacle fusion perception and semantic classification to determine the danger level of multiple obstacles in the environment in which the target vehicle is located. Based on the perceived data, a path planning is performed using a preset improved A* algorithm. During the node expansion process of the improved A* algorithm, the dynamic heuristic cost is determined based on the dynamic parameters corresponding to the danger level and the distance between the nodes to be processed. The improved actual cost is determined based on the path length and turning angle values ​​between the nodes to be processed. For the path inflection points in the nodes to be processed, the potential field constraint cost corresponding to the path inflection point is determined based on the repulsive force exerted by the obstacles within the visible range on the path inflection point. Based on the improved actual value and the dynamic heuristic value corresponding to the node to be processed, and the potential field constraint value corresponding to the path inflection point, the nodes to be processed are traversed to determine the parent node and the target node in each node to be processed. By tracing back from the target node to its parent node, the target driving path of the target vehicle in its environment is determined.

[0006] Furthermore, the step of using the lidar and camera equipment installed on the target vehicle to collect perception data under a unified spatiotemporal reference corresponding to the environment where the target vehicle is located includes: Spatiotemporal registration is performed on the lidar and camera equipment set on the target vehicle to establish the transformation relationship between the radar coordinate system corresponding to the lidar and the pixel coordinate system corresponding to the camera equipment. The raw perception data corresponding to the environment in which the target vehicle is located is collected using the lidar and camera equipment installed on the target vehicle. Based on the transformation relationship, the original sensing data is transformed to obtain sensing data under a unified spatiotemporal reference corresponding to the environment where the target vehicle is located.

[0007] Furthermore, the step of performing obstacle fusion perception and semantic classification on the perceived data to determine the danger levels corresponding to multiple obstacles in the environment where the target vehicle is located includes: Based on the radar perception data in the perception data, the lidar is used to determine the obstacle parameters corresponding to multiple obstacles in the environment where the target vehicle is located; The target detection model set in the camera device is used to identify multiple obstacles in the image perception data of the perception data, and the semantic category corresponding to each obstacle is determined; The obstacle parameters are fused and matched with the semantic category using multiple sensors to determine the location category information corresponding to each obstacle. Based on the location category information, the danger level of each obstacle is determined using a preset danger level classification rule.

[0008] Furthermore, determining the dynamic heuristic value based on the dynamic parameters corresponding to the danger level and the distance between the nodes to be processed includes: In the nodes to be processed, a plurality of adjacent nodes to be processed are determined, and the node distance between each node to be processed and each of the adjacent nodes to be processed is determined. Based on the danger level corresponding to each of the adjacent nodes to be processed, determine the dynamic parameters corresponding to each of the adjacent nodes to be processed. Based on the distance between nodes and the dynamic parameters, the dynamic heuristic value corresponding to each node to be processed is determined.

[0009] Furthermore, determining the actual cost of improvement based on the path length and turning angle values ​​corresponding to the nodes to be processed includes: Determine the sum of the actual distances from each node to be processed to the starting node, and use the sum of the actual distances to determine the cumulative cost of the path length corresponding to each node to be processed; Based on the angle between the direction from the parent node to the node to be processed and the direction from the node to be processed to each of the adjacent nodes to be processed, determine the path turning angle value corresponding to each node to be processed; Based on the path turning angle value and the preset turning cost coefficient, determine the additional turning angle cost corresponding to each node to be processed; The sum of the cumulative cost of the path length and the additional cost of the turning angle is determined as the improved actual cost value corresponding to each node to be processed.

[0010] Furthermore, for the path inflection points in the nodes to be processed, determining the potential field constraint cost corresponding to the path inflection point based on the repulsive force exerted by obstacles within the visible range on the path inflection point includes: For each path inflection point in the node to be processed, the straight-line Euclidean distance and repulsive force direction unit vector between each path inflection point and each obstacle within the visible range are determined, and the straight-line Euclidean distance is compared with a preset repulsive force distance threshold. When the straight-line Euclidean distance is less than or equal to the repulsive distance threshold, the repulsive force value of each obstacle in the visible range to each path inflection point is determined based on the straight-line Euclidean distance, the repulsive distance threshold, and the repulsive direction unit vector. When the straight-line Euclidean distance is greater than the repulsive distance threshold, the preset value is determined as the repulsive force value of each obstacle within the visible range on each path inflection point; Based on the repulsive force value, the total repulsive resultant force value corresponding to each path inflection point is determined, and the absolute value corresponding to the total repulsive resultant force value is determined as the potential field constraint cost value corresponding to the path inflection point.

[0011] Furthermore, the step of traversing the nodes to be processed based on the improved actual value and the dynamic heuristic value corresponding to the node to be processed, and the potential field constraint value corresponding to the path inflection point, to determine the parent node and the target node among the nodes to be processed for each node to be processed, includes: For a node to be processed that is not a path inflection point, the sum of the improved actual cost value and the dynamic heuristic cost value is determined as the target cost value corresponding to the node to be processed. For a node to be processed that is a path inflection point, the sum of the improved actual cost value, the dynamic heuristic cost value, and the potential field constraint cost value is determined as the target cost value corresponding to the node to be processed. Traverse the nodes to be processed, take the node to be processed corresponding to the first target value as the current node to be processed, and determine multiple neighboring nodes adjacent to the current node to be processed; wherein, when there are multiple nodes to be processed corresponding to the first target value, take the node to be processed corresponding to the first path turning angle value as the current node to be processed. For each of the neighboring nodes, if the neighboring node is an infeasible region node or has already been expanded, then no processing is performed on the neighboring node. If the neighboring node is not the node to be expanded, then the neighboring node is determined as the node to be expanded, and the current node to be processed is determined as the parent node corresponding to the neighboring node, and the target cost value corresponding to the neighboring node is determined. If the neighboring node is the node to be expanded and the actual improvement cost corresponding to the neighboring node is greater than the actual improvement cost corresponding to the current node to be processed, then the parent node corresponding to the neighboring node is determined as the current node to be processed. Repeatedly traverse the nodes to be processed until the target node is found.

[0012] This application embodiment also provides a vehicle path planning device based on LiDAR and a camera, the vehicle path planning device comprising: The semantic classification module is used to collect perception data under a unified spatiotemporal reference corresponding to the environment in which the target vehicle is located using the lidar and camera equipment installed on the target vehicle, and to perform obstacle fusion perception and semantic classification on the perception data to determine the danger level of multiple obstacles in the environment in which the target vehicle is located. The cost calculation module is used to perform path planning based on the perceived data using a preset improved A* algorithm. During the node expansion process of the improved A* algorithm, it determines the dynamic heuristic cost value based on the dynamic parameters corresponding to the danger level and the distance between the nodes to be processed, and determines the improved actual cost value based on the path length value and turning angle value between the nodes to be processed. For the path inflection points in the nodes to be processed, it determines the potential field constraint cost value corresponding to the path inflection point based on the repulsive force exerted by the obstacles within the visible range on the path inflection point. The node search module is used to traverse the nodes to be processed based on the improved actual value and the dynamic heuristic value corresponding to the node to be processed and the potential field constraint value corresponding to the path inflection point, so as to determine the parent node and the target node in the nodes to be processed for each node to be processed. The path generation module is used to trace back from the target node to the parent node level by level to determine the target driving path of the target vehicle in the environment.

[0013] This application embodiment also provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the vehicle path planning method based on LiDAR and camera as described above.

[0014] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the vehicle path planning method based on LiDAR and camera as described above.

[0015] This application provides a vehicle path planning method and apparatus based on LiDAR and cameras. The method includes: collecting perception data under a unified spatiotemporal reference corresponding to the environment where the target vehicle is located using LiDAR and camera equipment installed on the target vehicle; performing obstacle fusion perception and semantic classification on the perception data to determine the danger levels corresponding to multiple obstacles in the environment where the target vehicle is located; performing path planning based on the perception data using a preset improved A* algorithm; during the node expansion process of the improved A* algorithm, determining the dynamic heuristic value based on the dynamic parameters corresponding to the danger levels and the distance between nodes to be processed, and based on the distance between nodes to be processed... The path length and turning angle values ​​between points are used to determine the actual improvement cost. For the path inflection points in the nodes to be processed, the potential field constraint cost corresponding to the path inflection point is determined based on the repulsive force exerted by obstacles within the visible range on the path inflection point. Based on the actual improvement cost, the dynamic heuristic cost, and the potential field constraint cost corresponding to the path inflection point, the nodes to be processed are traversed to determine the parent node and the target node in each node to be processed. The target driving path of the target vehicle in the environment is determined by tracing back from the target node to the parent node.

[0016] Compared to the traditional A* algorithm in existing technologies used in actual vehicle applications, this improved A* algorithm uses the fusion perception of LiDAR and cameras to obtain the semantic category and spatial location of obstacles. Based on the semantic category of the obstacle, the danger level of the obstacle is determined. The cost function of the improved A* algorithm introduces the dynamic heuristic cost corresponding to the danger level, the additional cost of turning angle, and the cost of potential field constraint. Then, the improved A* algorithm is used to determine the target driving path of the vehicle. This effectively solves the technical defects of the traditional A* algorithm, such as single perception information, poor path smoothness, weak adaptability to dynamic environment, and the inability to balance search efficiency and planning accuracy. It realizes safe, efficient, and smooth path planning for vehicles in complex dynamic environments, significantly improves the reliability and adaptability of path planning, and fits various practical application scenarios of autonomous vehicles.

[0017] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a vehicle path planning method based on LiDAR and camera provided in an embodiment of this application; Figure 2 This is an example diagram of a vehicle driving path planning test provided in an embodiment of this application; Figure 3 A schematic diagram of a vehicle path planning device based on lidar and camera provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. Based on the embodiments of this application, every other embodiment obtained by those skilled in the art without inventive effort falls within the scope of protection of this application.

[0021] Research has shown that the A* algorithm is a heuristic search algorithm widely used in path planning. It combines the advantages of Dijkstra's algorithm and best-first search, selecting the optimal path by comprehensively considering the true cost (G value) from the starting point to the current node and the estimated cost (H value) from the current node to the target point. Due to its efficiency and accuracy, the A* algorithm performs particularly well in solving the shortest path in static road networks. However, with the increasing complexity of application scenarios, especially in the actual vehicle-side application of the traditional A* algorithm, the following limitations exist.

[0022] The system suffers from several drawbacks. First, it relies on limited and inadequate perception information. Path planning is based solely on grid maps, treating dynamic high-risk obstacles like pedestrians and vehicles indiscriminately compared to static low-risk obstacles like walls. This leads to planned paths frequently encountering high-risk obstacles, posing a collision hazard. Second, it exhibits poor path smoothness and adaptability. During node expansion, path length is used as the primary cost metric, resulting in numerous sharp turns and redundant inflection points. This does not align with vehicle dynamics, leading to poor vehicle stability and low path tracking accuracy. Third, it lacks dynamic adaptability and foresight. The static planning model focuses only on obstacles within the neighborhood, neglecting visible obstacles outside the neighborhood. This results in path inflection points frequently encountering obstacles, increasing the risk of collisions. Fourth, the fixed cost function fails to balance efficiency and accuracy. The heuristic cost uses fixed weights, leading to insufficient search accuracy in complex, high-risk environments and low efficiency in open, low-risk environments, making it unsuitable for all-weather, complex, and dynamic vehicle driving scenarios.

[0023] Based on this, this application provides a vehicle path planning method based on LiDAR and camera. It obtains the semantic category and spatial location of obstacles through the fusion perception of LiDAR and camera, determines the danger level of obstacles based on their semantic category, and introduces dynamic heuristic cost, turning angle additional cost, and potential field constraint cost into the cost function of the improved A* algorithm. Then, it uses the improved A* algorithm to determine the vehicle's target driving path. This effectively solves the technical defects of the traditional A* algorithm, such as single perception information, poor path smoothness, weak adaptability to dynamic environments, and the inability to balance search efficiency and planning accuracy. It achieves safe, efficient, and smooth path planning for vehicles in complex dynamic environments, significantly improving the reliability and adaptability of path planning, and fitting various practical application scenarios of autonomous vehicles.

[0024] Please see Figure 1 , Figure 1 This is a flowchart illustrating a vehicle path planning method based on LiDAR and a camera, provided as an embodiment of this application. Figure 1 As shown in the embodiments of this application, the vehicle path planning method based on LiDAR and camera is generally applicable to various vehicle scenarios requiring autonomous path planning, such as intelligent vehicles, autonomous vehicles, and unmanned delivery vehicles. It can achieve real-time and safe path planning in complex dynamic environments. The method includes: S101. Using the lidar and camera equipment installed on the target vehicle, perception data under a unified spatiotemporal reference corresponding to the environment where the target vehicle is located is collected, and obstacle fusion perception and semantic classification are performed on the perception data to determine the danger level corresponding to multiple obstacles in the environment where the target vehicle is located.

[0025] In this embodiment of the application, a lidar is installed on the target vehicle to collect three-dimensional point cloud data of the vehicle's surrounding environment. It is a sensor that can accurately obtain spatial physical information such as the distance, position, shape, and size of obstacles. A camera device is installed on the target vehicle and used in conjunction with the lidar to collect two-dimensional image data of the vehicle's surrounding environment. It is a sensor that can identify the semantic category of obstacles (such as pedestrians, vehicles, and static obstacles).

[0026] Here, the target vehicle refers to an autonomous vehicle, intelligent vehicle, etc., that requires path planning and is the subject of execution of the method described in this application embodiment.

[0027] In this embodiment of the application, a unified spatiotemporal reference refers to unifying the radar coordinate system of the lidar and the pixel coordinate system of the camera to eliminate the spatiotemporal deviation between the two, so that the point cloud data collected by the lidar and the image data collected by the camera are in the same spatiotemporal dimension, thus ensuring the accuracy of data fusion.

[0028] The perception data includes radar perception data (e.g., point cloud data) collected by lidar and image perception data (e.g., two-dimensional images) collected by camera, which are then spatiotemporally registered to form fused perception data under a unified spatiotemporal reference.

[0029] In this embodiment of the application, obstacle fusion perception refers to matching and fusing the spatial physical information of obstacles acquired by lidar with the semantic information of obstacles acquired by camera to achieve a one-to-one correspondence between "obstacle location and obstacle category" and make up for the defects of a single sensor; semantic classification refers to classifying obstacles in the perception data to distinguish different types of obstacles such as pedestrians, vehicles, and static obstacles (such as walls and curbs).

[0030] Here, the danger level is a risk level classified according to the semantic category of the obstacle. It is used to dynamically adjust the accuracy and efficiency of path planning. High risk corresponds to high-precision search, and low risk corresponds to high-efficiency search.

[0031] In this step, firstly, a LiDAR and camera device are installed on the target vehicle. The two devices are spatiotemporally registered to establish a transformation relationship between the radar coordinate system and the pixel coordinate system, eliminating spatiotemporal deviations. Then, the two devices are used to collect raw perception data of the vehicle's surrounding environment (the LiDAR collects point cloud data, and the camera collects image data). Based on the transformation relationship, the raw perception data is converted into perception data under a unified spatiotemporal reference. Finally, the perception data is fused and semantically classified. The spatial parameters of obstacles are obtained through the LiDAR, and the semantic categories of obstacles are obtained through the camera. The two are matched and bound, and then the danger level of each obstacle is classified according to preset rules.

[0032] This step achieves triple recognition of obstacles' "location, category, and danger level" by acquiring accurate perception information of the vehicle's surrounding environment. It makes up for the lack of semantic information about obstacles in the traditional A* algorithm, provides core input for the subsequent cost function calculation of the A* algorithm, and lays the foundation for balancing the safety and efficiency of path planning.

[0033] This eliminates the spatiotemporal discrepancies between lidar and cameras, enabling precise fusion of multi-sensor data and addressing the shortcomings of single sensors that "know the location but not the category" (lidar) or "know the category but not the distance" (camera). Through semantic classification and hazard level division, path planning can distinguish obstacle risks, providing a basis for subsequent dynamic adjustment of search accuracy and efficiency, and improving the safety of path planning.

[0034] In one possible implementation of this application, in specific implementation, the step S101 of collecting perception data under a unified spatiotemporal reference corresponding to the environment where the target vehicle is located using the lidar and camera equipment installed on the target vehicle may include: S1011. Perform spatiotemporal registration on the lidar and camera equipment set on the target vehicle to establish the transformation relationship between the radar coordinate system corresponding to the lidar and the pixel coordinate system corresponding to the camera equipment.

[0035] In the embodiments of this application, spatiotemporal registration refers to calibrating the lidar and camera equipment, establishing the transformation relationship between the radar coordinate system and the pixel coordinate system, eliminating the spatiotemporal deviation between the two, and ensuring the consistency of the perceived data.

[0036] Here, the radar coordinate system is a three-dimensional coordinate system established with the lidar as the origin, used to describe the spatial location of the point cloud data acquired by the lidar; the pixel coordinate system is a two-dimensional coordinate system established with the upper left corner of the camera image as the origin, used to describe the position of the pixels in the image acquired by the camera.

[0037] The transformation relationship refers to the coordinate transformation formula between the radar coordinate system and the pixel coordinate system, which is used to transform the point cloud data of the lidar to the pixel coordinate system, or vice versa, to achieve data fusion.

[0038] In this step, spatiotemporal registration is the core step to eliminate the spatiotemporal deviation between the lidar and the camera equipment. Its core is to establish a mathematical model for coordinate transformation between the lidar coordinate system and the pixel coordinate system, so that the point cloud data collected by the lidar can be accurately matched with the image data collected by the camera. The establishment of the transformation relationship is based on the camera intrinsic parameters, the rotation matrix and translation matrix of the lidar and the camera. The relevant parameters are determined through calibration experiments, and finally a unified coordinate transformation formula is obtained.

[0039] In this embodiment of the application, the coordinate transformation formula corresponding to the transformation relationship between the radar coordinate system corresponding to the lidar and the pixel coordinate system corresponding to the camera device is as follows.

[0040] .

[0041] in, Represents pixel coordinates, ; Indicates radar coordinates, ; This represents the intrinsic parameter matrix of the camera device; and These represent the rotation and translation matrices from the radar coordinate system to the pixel coordinate system, respectively. This represents the normalization coefficient.

[0042] S1012. Use the lidar and camera equipment installed on the target vehicle to collect raw perception data corresponding to the environment where the target vehicle is located.

[0043] Here, raw sensing data refers to raw point cloud data and raw image data directly collected by LiDAR and camera without spatiotemporal registration. It has spatiotemporal deviations and cannot be directly used for fusion sensing.

[0044] In this step, after completing the spatiotemporal registration, the lidar and camera equipment are activated to simultaneously collect raw perception data of the surrounding environment of the target vehicle.

[0045] Among them, the lidar collects three-dimensional point cloud data to obtain spatial information of obstacles within a first preset range around the vehicle; the camera collects two-dimensional image data to obtain environmental image information within a second preset range around the vehicle, thus obtaining raw perception data.

[0046] S1013. Based on the transformation relationship, the original perception data is transformed to obtain perception data under a unified spatiotemporal reference corresponding to the environment where the target vehicle is located.

[0047] In this step, based on the established coordinate transformation relationship, the raw point cloud data collected by the lidar is transformed into the pixel coordinate system of the camera, and the raw image data collected by the camera is associated with the lidar coordinate system, thereby achieving spatiotemporal alignment of the two raw sensing data.

[0048] Here, the aligned data is the perception data under a unified spatiotemporal reference, including radar perception data (aligned point cloud data) and image perception data (aligned image data), which can be directly used for subsequent obstacle fusion perception and semantic classification.

[0049] In one possible implementation of this application, in specific implementation, the step S101 of performing obstacle fusion perception and semantic classification on the perceived data to determine the danger levels corresponding to multiple obstacles in the environment where the target vehicle is located may include: S1014. Based on the radar perception data in the perception data, use the lidar to determine the obstacle parameters corresponding to multiple obstacles in the environment where the target vehicle is located.

[0050] Here, obstacle parameters refer to the spatial physical parameters related to obstacles collected by lidar, including the obstacle's three-dimensional coordinates, distance, size, outline, and position in the grid map.

[0051] In this embodiment of the application, the radar sensing data is lidar point cloud data under a unified spatiotemporal reference. By clustering and segmenting the point cloud data, removing ground points and noise points, the point cloud clusters corresponding to obstacles are extracted, and then the spatial physical parameters of each obstacle cluster, i.e. obstacle parameters, are calculated. These parameters can accurately describe the spatial location and shape of the obstacle, providing a spatial basis for subsequent fusion matching.

[0052] S1015. Using the target detection model set in the camera device, identify multiple obstacles in the image perception data of the perception data, and determine the semantic category corresponding to each obstacle.

[0053] In this embodiment of the application, the image perception data is camera image data under a unified spatiotemporal reference. The image data is used to perform inference and recognition using a pre-trained target detection model to detect all obstacles in the image and perform semantic classification on each obstacle to determine its category (such as pedestrian, vehicle, static obstacle), providing a category basis for subsequent hazard level classification.

[0054] Among them, the target detection model can be either the YOLO model or the Mask R-CNN model, which has high recognition accuracy and real-time performance and is suitable for vehicle applications.

[0055] For example, a pre-trained YOLOv8 object detection model is selected to perform inference and recognition on camera image data under a unified spatiotemporal benchmark. The confidence threshold is set to 0.8 and the IOU threshold is set to 0.5 to detect obstacles in the image. The semantic categories of obstacles are divided into three categories: pedestrians (dynamic high risk), vehicles (dynamic medium risk), and static obstacles (such as walls and curbs, static low risk). For example, three obstacles are identified in the image: obstacle 1 (semantic category: pedestrian), obstacle 2 (semantic category: vehicle), and obstacle 3 (semantic category: static obstacle - wall).

[0056] S1016. Perform multi-sensor fusion matching between the obstacle parameters and the semantic category to determine the location category information corresponding to each obstacle.

[0057] Here, multi-sensor fusion matching refers to accurately matching the obstacle parameters acquired by LiDAR with the obstacle semantic categories acquired by the camera based on the transformation relationship after spatiotemporal registration, so as to achieve the binding of the "position pre-category" of each obstacle.

[0058] Among them, location category information refers to the comprehensive information of "spatial location and semantic category" corresponding to each obstacle, which is the core basis for determining the danger level of obstacles.

[0059] In this step, based on a unified spatiotemporal reference, obstacle parameters (spatial location) are accurately matched with obstacle semantic categories. Through coordinate association, the "spatial location and semantic category" of each obstacle are bound together to obtain location category information. This can eliminate the recognition error of a single sensor and ensure that the location and category of the obstacle correspond correctly.

[0060] S1017. Based on the location category information, determine the danger level corresponding to each obstacle using a preset danger level classification rule.

[0061] In this embodiment of the application, the hazard level classification rule is a preset rule that classifies hazard levels according to the semantic category of obstacles, and is used to unify the hazard level assignment standard.

[0062] In this step, the danger level is classified according to the semantic category of the obstacle. High-risk obstacles correspond to low danger level values, and low-risk obstacles correspond to high danger level values, which are used for the calculation of subsequent dynamic parameters. Based on the semantic category in the location category information, and in accordance with the preset danger level classification rules, a danger level is assigned to each obstacle.

[0063] For example, the preset hazard level classification rule is as follows: the hazard level D of the semantic category of pedestrian (dynamic high risk) is 1; the hazard level D of the semantic category of vehicle (dynamic medium risk) is 2; and the hazard level D of the semantic category of static obstacle (static low risk) is 3. According to this rule, combined with the location category information, the hazard levels of the three obstacles are determined as follows: obstacle 1 (pedestrian) has a D of 1, obstacle 2 (vehicle) has a D of 2, and obstacle 3 (static obstacle) has a D of 3.

[0064] S102. Based on the perceived data, a path planning is performed using a preset improved A* algorithm. During the node expansion process of the improved A* algorithm, the dynamic heuristic cost is determined based on the dynamic parameters corresponding to the danger level and the distance between the nodes to be processed. The improved actual cost is determined based on the path length and turning angle values ​​between the nodes to be processed. For the path inflection points in the nodes to be processed, the potential field constraint cost corresponding to the path inflection point is determined based on the repulsive force exerted by the obstacles within the visible range on the path inflection point.

[0065] In the embodiments of this application, the improved A* algorithm is based on the traditional A* algorithm. By reconstructing the cost function, introducing dynamic heuristic cost, improving the actual cost and potential field constraint cost, and optimizing the node expansion rules, it solves the defects of the traditional A* algorithm and is an algorithm adapted to the needs of vehicle-side path planning.

[0066] Among them, the nodes to be processed refer to the grid nodes in the improved A* algorithm that are in the process of path search and have not yet been determined to be the optimal path nodes, including nodes to be expanded and nodes that have been expanded.

[0067] In this embodiment, the dynamic heuristic cost is a component of the improved A* algorithm cost function, determined based on dynamic parameters corresponding to obstacle danger levels and the distance between nodes to be processed, which can dynamically balance the accuracy and efficiency of path search; the improved actual cost is a component of the improved A* algorithm cost function, determined based on the path length and turning angle values ​​between nodes to be processed, used to penalize large-angle turns and improve path smoothness; the potential field constraint cost is a component of the improved A* algorithm cost function, determined based on the repulsive force of obstacles on path inflection points, used to constrain the position of path inflection points and improve path safety.

[0068] Here, a path inflection point refers to a node in the path where the direction changes. It is a key factor affecting the smoothness of the path and the core object of the artificial potential field repulsive force constraint. The repulsive force is a pushing force exerted by the obstacle on the path inflection point based on the artificial potential field method. Its direction is away from the obstacle and is used to prevent the path inflection point from getting close to the obstacle, thereby improving the path safety.

[0069] In this step, based on the perception data under a unified spatiotemporal reference, the preset improved A-Star algorithm is launched for path planning. During the node expansion process of the algorithm, the three components of the cost function are calculated in three steps: First, dynamic parameters are determined according to the danger level of obstacles, and the dynamic heuristic cost is calculated in combination with the distance between the nodes to be processed, so as to realize the dynamic adjustment of the heuristic cost; Second, the actual cost of improvement is calculated based on the sum of the path lengths (cumulative cost of path length) and the turning angle of the path between the nodes to be processed, and the turning angle cost penalty for large-angle turns is introduced; Third, for the inflection points in the path, the repulsive force of each obstacle in the visible range to the inflection point is calculated, and the total repulsive resultant force is obtained by superimposing them. The absolute value of the total repulsive resultant force is used as the potential field constraint cost to constrain the inflection point away from the obstacle.

[0070] In this embodiment, the step reconstructs and improves the cost function of the A* algorithm, which solves the shortcomings of the traditional A* algorithm, such as a single cost function, inability to balance search accuracy and efficiency, poor path smoothness, and lack of obstacle-oriented constraints. Through the synergistic effect of the three cost components, the path planning can ensure both efficiency and accuracy, while also improving smoothness and safety.

[0071] In this way, the dynamic heuristic value achieves adaptive search with "high accuracy in high-risk areas and high efficiency in low-risk areas", balancing the accuracy and efficiency of path planning; the improved practical value penalizes large-angle turns, reduces sharp turns and redundant inflection points in the path, improves path smoothness, and adapts to vehicle driving dynamics; the potential field constraint value enables path inflection points to actively move away from obstacles, improving the safety and foresight of the path and avoiding the collision risk caused by inflection points being close to obstacles.

[0072] In one possible implementation of this application, in specific implementation, the step of determining the dynamic heuristic value based on the dynamic parameters corresponding to the danger level and the distance between the nodes to be processed in step S102 may include: S10211. Among the nodes to be processed, determine a plurality of adjacent nodes to be processed that are adjacent to each node to be processed, and determine the node distance between each node to be processed and each of the adjacent nodes to be processed.

[0073] Here, the distance between nodes refers to the straight-line distance between two adjacent nodes to be processed.

[0074] In this step, the nodes to be processed are grid nodes in the improved A* algorithm. An 8-neighborhood search strategy is adopted to determine the 8 neighboring nodes (up, down, left, right, and four diagonal directions) for each node to be processed. The distance between nodes is calculated using Euclidean distance, which is the straight-line distance between the grid coordinates of two adjacent nodes to be processed, and serves as the basic parameter for dynamic heuristic cost calculation.

[0075] For example, the grid size of the raster map is set to 0.5m × 0.5m. The grid coordinates of a node A to be processed are (50, 10), and the grid coordinates of its 8 adjacent nodes to be processed are (49,9), (49,10), (49,11), (50,9), (50,11), (51,9), (51,10), (51,11). The Euclidean distance from node A to each adjacent node is calculated: the distance between horizontally / vertically adjacent nodes (such as (50,10) to (50,11)) is 0.5m, and the distance between diagonally adjacent nodes (such as (50,10) to (49,9)) is "0.5 × √2 ≈ 0.707m".

[0076] S10212. Based on the danger level corresponding to each of the adjacent nodes to be processed, determine the dynamic parameters corresponding to each of the adjacent nodes to be processed.

[0077] Here, the coefficients of the dynamic parameters that change dynamically with the obstacle danger level are used to adjust the weight of the dynamic heuristic cost value. The higher the danger level, the smaller the dynamic parameters and the higher the search accuracy.

[0078] In this step, the dynamic parameter is a coefficient that changes dynamically with the obstacle hazard level; the hazard level of the adjacent node to be processed is determined by the obstacle hazard level within the grid range where the node is located (if there are multiple obstacles within the range of the node, the D value with the lowest hazard level is taken, i.e., the highest risk); based on the hazard level, the dynamic parameter corresponding to the adjacent node to be processed is calculated by substituting it into the formula.

[0079] In this embodiment of the application, the dynamic parameters are determined by the following formula.

[0080] .

[0081] in, Indicates dynamic parameters; This represents the preset maximum dynamic parameter (system constant, calibrated according to the real-time and accuracy requirements of the vehicle). Indicates the level of danger.

[0082] For example, assume a preset maximum dynamic parameter If the adjacent node B (grid (50,11)) contains obstacle 1 (D=1), then the danger level of node B is D=1, and the corresponding dynamic parameter is 3. =3; If there is an obstacle (D=3) within the grid range of the adjacent node C (grid (45,8)), then the danger level of node C is D=3, and the corresponding dynamic parameter is... =1; If there is obstacle 2 (D=2) within the grid range of the adjacent node D (grid (55,12)), then the danger level of node D is D=2, and the corresponding dynamic parameter is... =1.5.

[0083] S10213. Based on the distance between nodes and the dynamic parameters, determine the dynamic heuristic value corresponding to each node to be processed.

[0084] In this embodiment of the application, the formula for calculating the dynamic heuristic cost value is: the formula for calculating the dynamic heuristic cost value is "e α·h(n) “”, where e is the natural constant (approximately equal to 2.718), α is the dynamic parameter corresponding to the adjacent node to be processed, and h(n) is the distance between the node to be processed and the adjacent node to be processed.

[0085] In this step, the distance between nodes and dynamic parameters are substituted into the formula to calculate the dynamic heuristic cost corresponding to each node to be processed.

[0086] In this embodiment, obstacles are semantically classified, and the parameters in the heuristic function are dynamically adjusted according to the danger level of the obstacles. This improves path search accuracy in complex obstacle environments and enhances search efficiency in simple environments. By utilizing the fusion perception of LiDAR and cameras, obstacles are semantically classified (e.g., pedestrians, vehicles, walls, etc.) and assigned different danger levels. Dynamic parameters are dynamically adjusted according to the danger level of the obstacle; higher danger levels result in lower dynamic parameters for more precise path search, while lower danger levels result in higher dynamic parameters for faster search.

[0087] In one possible implementation of this application, in specific implementation, the step of determining the actual improvement cost based on the path length value and turning angle value corresponding to the nodes to be processed in step S102 may include: S10221. Determine the sum of the actual distances from each node to be processed to the starting node, and use the sum of the actual distances to determine the cumulative cost of the path length corresponding to each node to be processed.

[0088] In this embodiment, the cumulative cost of path length refers to the total actual distance from the starting node to the current node to be processed, which is a fundamental component for improving the actual cost value.

[0089] In this step, the cumulative cost of the path length is calculated by accumulating the distances between nodes. For the starting node, its cumulative cost of the path length is 0. For other nodes to be processed, its cumulative cost of the path length is the sum of the cumulative cost of the path length of its parent node and the distance between the current node and its parent node, which is used to reflect the actual path cost from the starting point to the current node.

[0090] S10222. Based on the angle between the direction from the parent node to the node to be processed and the direction from the node to be processed to each adjacent node to be processed, determine the path turning angle value corresponding to each node to be processed.

[0091] In this embodiment, the path turning angle value refers to the angle between the driving direction from the parent node of the current node to be processed to the node to be processed and the driving direction from the node to be processed to the adjacent node to be processed, and is used to calculate the additional cost of turning angle.

[0092] Here, the driving direction is represented by a vector. The direction vector is calculated using the grid coordinates of two nodes, and then the angle between the two direction vectors is calculated, which is the turning angle value of the path. The angle ranges from 0° (no turn) to 180° (reverse turn). The larger the angle, the larger the turning range.

[0093] S10223. Based on the turning angle value of the path and the preset turning cost coefficient, determine the additional turning angle cost corresponding to each node to be processed.

[0094] Here, the turning cost coefficient is a preset coefficient used to adjust the weight of the additional cost of turning angle, and it is calibrated according to the requirements of vehicle turning energy consumption and driving stability; the additional cost of turning angle is an additional cost determined based on the turning angle value of the path and the turning cost coefficient. The larger the turning angle, the higher the additional cost, which is used to penalize large-angle turns.

[0095] In this embodiment of the application, the additional cost of turning angle is determined by the following formula.

[0096] .

[0097] in, Add a cost to the turning angle corresponding to each node to be processed; For the current node To neighboring nodes The Euclidean distance is the cost of travel; for point to direction and point to The angle between the directions, for Adjacent nodes; and The turning cost coefficient is a preset value, which is determined based on the energy consumption ratio and time ratio of turning in actual applications.

[0098] Thus, the total surcharge value obtained from the new calculation formula... Compare Added "(1+μ Therefore, when expanding nodes, nodes with fewer turning points will be selected first.

[0099] S10224. The sum of the cumulative cost of the path length and the additional cost of the turning angle is determined as the improved actual cost value corresponding to each node to be processed.

[0100] In this embodiment, the improved actual cost is the sum of the cumulative cost of the path length and the additional cost of the turning angle, and its calculation formula is "g * (n)=g(n)+t(n)”; This index takes into account both the actual length of the path and the influence of the turning angle, and can effectively penalize large-angle turns and improve the smoothness of the path.

[0101] Here, the traditional A* algorithm only calculates the path length cost in its actual cost update. To reduce the impact of the number of turns in path planning, this application proposes an optimization model that comprehensively considers path length and turning angle. In the process of expanding sub-nodes in path planning, the turning angle is used as an additional evaluation index for selecting nodes to reduce path complexity and improve path smoothness.

[0102] In one possible implementation of this application, in specific implementation, step S102, for the path inflection point in the node to be processed, the step of determining the potential field constraint value corresponding to the path inflection point based on the repulsive force exerted on the path inflection point by obstacles within the visible range, may include: S10231. For the path inflection points in the node to be processed, determine the straight-line Euclidean distance and repulsive force direction unit vector between each path inflection point and each obstacle within the visible range, and compare the straight-line Euclidean distance with a preset repulsive force distance threshold.

[0103] Here, the straight-line Euclidean distance refers to the straight-line distance between the path inflection point and the obstacle, which is the core parameter for calculating the repulsive force; the repulsive force direction unit vector is the repulsive force direction unit vector of the obstacle to the path inflection point in the artificial potential field. It has both direction (from the obstacle to the inflection point, i.e., the direction away from the obstacle) and fixed magnitude (magnitude of 1). It is only used to determine the repulsive force direction and does not change the repulsive force magnitude.

[0104] In this step, firstly, path inflection points (nodes where the direction changes) are selected from the nodes to be processed; then, the visible range of each path inflection point is determined (based on the vehicle's field of vision and obstacle occlusion, excluding occluded obstacles); for each path inflection point, the straight-line Euclidean distance between it and each obstacle within the visible range is calculated, and the unit vector of the repulsive force direction is also calculated; finally, the straight-line Euclidean distance is compared with a preset repulsive force distance threshold to determine whether the repulsive force needs to be calculated.

[0105] In this embodiment of the application, the straight-line Euclidean distance between each path inflection point and each obstacle within the visible range is determined by the following formula.

[0106] .

[0107] in, Indicate each path inflection point With each obstacle within the line of sight The straight-line Euclidean distance between them; k represents the current position, and is the position of the inflection point only when θ is greater than zero. , Indicates the grid position of the obstacle.

[0108] In this embodiment of the application, the repulsive force direction unit vector between each path inflection point and each obstacle within the visible range is determined by the following formula.

[0109] .

[0110] in, For each path inflection point With each obstacle within the line of sight The repulsive force between them is a unit vector. , Indicates the grid position of obstacles; , is the unit basis vector for the X and Y axes.

[0111] S10232. When the straight-line Euclidean distance is less than or equal to the repulsive distance threshold, the repulsive force value of each obstacle within the visible range to each path inflection point is determined based on the straight-line Euclidean distance, the repulsive distance threshold, and the repulsive direction unit vector.

[0112] Here, the repulsion distance threshold is a preset effective range threshold for repulsion. When the straight-line Euclidean distance between the path inflection point and the obstacle is less than or equal to this threshold, repulsion is generated; when it is greater than this threshold, the repulsion is 0. The repulsion force value is the magnitude of the repulsion force exerted by a single obstacle on a single path inflection point, which is determined by the straight-line Euclidean distance, the repulsion distance threshold, and the unit vector of the repulsion direction.

[0113] In this embodiment of the application, the repulsive force value of each obstacle on each path inflection point within the visible range is determined by the following formula.

[0114] .

[0115] Here, each obstacle within the visible range is represented. For each path inflection point The repulsive force value; For each path inflection point With each obstacle within the line of sight The repulsive force between them is a unit vector. Indicate each path inflection point With each obstacle within the line of sight The straight-line Euclidean distance between them; Indicates the repulsive distance threshold; The preset repulsion index (calibrated according to the repulsion strength requirements).

[0116] Here, the logic behind the formula for determining the repulsive force value is: the closer the path inflection point is to the obstacle, the greater the repulsive force, and the direction is away from the obstacle.

[0117] S10233. When the straight-line Euclidean distance is greater than the repulsive distance threshold, the preset value is determined as the repulsive force value of each obstacle within the visible range on each path inflection point.

[0118] In this step, when the Euclidean distance between the straight line is greater than the repulsion distance threshold, it indicates that the distance between the path inflection point and the obstacle exceeds the effective range of the repulsion force. The obstacle cannot effectively repel the path inflection point. At this time, there is no need to apply repulsion constraints to avoid unnecessary constraints increasing the computational load of the algorithm, while ensuring the efficiency of path planning.

[0119] In this embodiment of the application, the preset value 0 is determined as the repulsive force value in this case, that is, the obstacle has no repulsive force effect on the path inflection point.

[0120] S10234. Based on the repulsive force value, determine the total repulsive resultant force value corresponding to each path inflection point, and determine the absolute value corresponding to the total repulsive resultant force value as the potential field constraint cost value corresponding to the path inflection point.

[0121] Here, the total repulsive resultant force is the sum of the repulsive forces of all visible obstacles at a single path inflection point, used to determine the cost of potential field constraints.

[0122] In this embodiment, the total repulsive force value is the vector summation of the repulsive forces of all obstacles within the visible range at a single path inflection point. The repulsive force value (vector) of each obstacle at the path inflection point needs to be summed to obtain the total repulsive force value. Since the potential field constraint cost only needs to reflect the constraint strength of the repulsive force and is independent of the direction, the absolute value of the total repulsive force value is used as the potential field constraint cost value corresponding to the path inflection point. The larger the value, the stronger the repulsive constraint at the path inflection point, and the more it needs to stay away from the obstacles, thereby improving the safety of the path.

[0123] Here, in order to improve the foresight of path planning and make obstacles outside the neighborhood affect path points, an artificial potential field is introduced to constrain path points near obstacles, thereby improving the smoothness and safety of the path.

[0124] The basic principle of the artificial potential field method is to treat obstacles as a source of repulsive force. The magnitude of the repulsive force on the path points gradually decreases as the distance between the path point and the obstacle increases, thus moving the inflection points near the obstacle to the place where the resultant force is minimal. In the hybrid method, the repulsive force field generated by the obstacle is used to apply a repulsive force to the surrounding path points, causing the path points to move away from the obstacle.

[0125] It should be noted that, firstly, in order to reduce the amount of computation, the influence range of the obstacle is only effective within a local dynamic window, and this range also needs to take into account the size and dynamic characteristics of the vehicle, and should not be too small; secondly, in order not to increase the length of the path, the influence of the repulsive field on the inflection point is much greater than that on other ordinary nodes.

[0126] In practical implementation, regarding the collision problem, considering that applying repulsive forces to all obstacle grid points within the window would increase computational overhead and generate redundant force fields, repulsive forces are only applied to obstacles within the visible range, while no force is applied to occluded inner-layer obstacles. This method ensures that turning points are far from obstacles that may collide, while avoiding unnecessary computational overhead and redundant force field interference.

[0127] S103. Based on the improved actual value and the dynamic heuristic value corresponding to the node to be processed, and the potential field constraint value corresponding to the path inflection point, traverse the nodes to be processed to determine the parent node and the target node among the nodes to be processed for each node to be processed.

[0128] Here, the parent node is the predecessor node that guides the current node to be processed to join the optimal path during the path search process, and is used for subsequent path backtracking to determine the complete driving path; the target node refers to the end point of the path planning, that is, the location that the target vehicle needs to reach.

[0129] In this step, based on the type of node to be processed (path inflection point / non-path inflection point), the target cost value of each node to be processed is calculated (for non-inflection points, it is the sum of the actual improvement cost value and the dynamic heuristic cost value; for inflection points, it is the sum of all three). An 8-neighborhood search strategy is used to traverse all nodes to be processed. Each time, the node with the smallest target cost value is selected as the current node to be processed. Its neighboring nodes are then filtered and processed to determine the parent node (current node to be processed) of each neighboring node, and the list of nodes to be expanded and the list of expanded nodes are updated. The above process is repeated until the target node (path endpoint) is found.

[0130] In this embodiment of the application, this step filters the optimal node by target cost value, reasonably determines the parent node of each node to be processed, ensures the optimality and efficiency of path search, and finally finds the optimal feasible path from the starting node to the target node, while avoiding the repeated processing of invalid nodes and reducing the amount of computation.

[0131] In this way, efficient traversal and filtering of nodes to be processed are achieved, ensuring that each selected node is the current optimal node, thus guaranteeing the optimality of the path. The reasonable determination of parent nodes provides a clear logical basis for subsequent path backtracking. The node filtering rules avoid the processing of infeasible and duplicate nodes, reduce the computational load of the algorithm, and meet the real-time requirements of the vehicle.

[0132] In one possible implementation of this application, step S103 may include: S1031. For the node to be processed that is not a path inflection point, the sum of the improved actual cost value and the dynamic heuristic cost value is determined as the target cost value corresponding to the node to be processed. For the node to be processed that is a path inflection point, the sum of the improved actual cost value, the dynamic heuristic cost value and the potential field constraint cost value is determined as the target cost value corresponding to the node to be processed.

[0133] Here, the target cost is the core indicator for selecting the optimal node in the improved A* algorithm. The target cost of non-path inflection points is the sum of the actual improvement cost and the dynamic heuristic cost, while the target cost of path inflection points is the sum of all three.

[0134] In the embodiments of this application, non-path inflection points only serve the function of path passage, and their target value is composed of the improved actual value and the dynamic heuristic value. As a key position for the change of path direction, path inflection points need to bear additional repulsive constraint pressure from obstacles. Therefore, their target value needs to be superimposed with the potential field constraint value to ensure that the inflection point position is far away from danger.

[0135] In this way, by distinguishing node types to calculate target cost value, fine-grained control over the overall safety and smoothness of the path is achieved.

[0136] S1032. Traverse the nodes to be processed, take the node to be processed corresponding to the first target value as the current node to be processed, and determine multiple neighboring nodes adjacent to the current node to be processed.

[0137] When there are multiple nodes to be processed corresponding to the first target value, the node to be processed corresponding to the first path turning angle value is taken as the current node to be processed.

[0138] Here, a neighboring node refers to a node that is adjacent to the current node to be processed, and the neighboring nodes determined by the 8-neighborhood search strategy (i.e., nodes in the four diagonal directions above, below, left, right, and above the current node).

[0139] In this embodiment of the application, this step is the core startup step of the improved A* algorithm node expansion. The core purpose is to select the current optimal node to be processed as the current node, so as to lay the foundation for subsequent neighbor node processing and parent node determination.

[0140] In this step, firstly, all nodes to be processed are fully traversed to select the node with the smallest target value. This smallest target value is the first target value, and the corresponding node to be processed is the initial current node to be processed. Then, an 8-neighborhood search strategy (covering the current node's top, bottom, left, right, and four diagonal directions) is used to determine all neighboring nodes adjacent to the current node to be processed. These neighboring nodes are potential candidate nodes for subsequent path expansion.

[0141] Furthermore, if after traversal it is found that multiple nodes to be processed have the same target cost value as the first target cost value (i.e. multiple nodes have the same cost), then the turning angle value of the path is introduced as a secondary screening criterion. The node with the smallest turning angle value of the path (i.e. the first turning angle value of the path) is selected as the current node to be processed. This design can reduce sharp turns in the path from the source, further improve the smoothness of the path, conform to the dynamic characteristics of vehicle driving, and avoid the path non-smoothness problem caused by random point selection when the cost is the same in the traditional A* algorithm.

[0142] For example, suppose the nodes to be processed include node M (f=18.1832), node K (f=63.5296), node N (f=18.1832), and other nodes. After traversing all the nodes to be processed, the first target cost value is determined to be 18.1832, and the corresponding nodes to be processed are node M and node N (the two nodes have the same target cost value). At this time, it is necessary to further compare the path turning angle values ​​of the two nodes: the path turning angle value of node M is θ=30°, the path turning angle value of node N is θ=60°, the first path turning angle value is 30°, so node M is selected as the current node to be processed; then, using the 8-neighborhood search strategy, the neighboring nodes of node M are determined to be N1, N2, N3, N4, N5, N6, N7, and N8 (adjacent nodes to be processed in 8 directions). These neighboring nodes are used as candidate nodes for subsequent path expansion and enter the next processing stage.

[0143] S1033. For each of the neighboring nodes, if the neighboring node is an infeasible region node or has been expanded, then no processing is performed on the neighboring node.

[0144] Here, infeasible area nodes refer to unprocessed nodes in the grid map that are occupied by obstacles and cannot be passed by vehicles; expanded processing nodes refer to unprocessed nodes that have been processed and whose optimal parent node has been determined, and are stored in the Closed list.

[0145] In this embodiment, this step is the neighbor node screening process. Its core purpose is to eliminate invalid neighbor nodes, reduce the computational load of the algorithm, avoid invalid nodes occupying computational resources, and ensure the efficiency and accuracy of node expansion.

[0146] Among them, nodes in infeasible areas would cause collision risks in path planning if processed, so they are directly removed; nodes that have been expanded and processed refer to nodes that have been included in the expanded node list (Closed list), whose optimal parent node has been determined and processed. These nodes do not need to be processed repeatedly, avoiding redundant calculations in the algorithm and ensuring the orderly expansion of nodes.

[0147] For example, suppose node M has neighboring nodes N1 to N8. Each neighboring node is filtered: the grid where neighboring node N1 is located is occupied by an obstacle, belonging to an infeasible area node, and no processing is done on N1; neighboring node N2 has been included in the expanded node list (Closed list), belonging to a node that has been expanded and processed, and no processing is done on N2; neighboring nodes N3 to N8 are neither infeasible area nodes nor expanded and are valid candidate nodes, entering the subsequent processing stage to ensure that node expansion is only performed on valid and feasible nodes.

[0148] S1034. If the neighboring node is not the node to be expanded, then the neighboring node is determined as the node to be expanded, and the current node to be processed is determined as the parent node corresponding to the neighboring node, and the target value corresponding to the neighboring node is determined.

[0149] Here, nodes to be expanded refer to nodes that have been discovered but not yet processed, and are stored in the Open list.

[0150] In this embodiment of the application, this step is the addition and parameter configuration of the node to be expanded. The core purpose is to include the selected effective neighboring nodes into the list of nodes to be expanded (Open list), clarify their parent-child node relationships, and complete the calculation of the target cost value, so as to lay the foundation for subsequent node expansion and screening.

[0151] Among them, "not a node to be expanded" means that the neighboring node has not been included in the list of nodes to be expanded (Open list) and is a valid candidate node discovered for the first time. After it is identified as a node to be expanded, its parent node is bound to the current node to be processed (to ensure the continuity of path backtracking). Then, according to the type of the neighboring node (path inflection point / non-path inflection point), its corresponding target value is determined according to the above target value calculation rules, and the initial configuration of the node is completed.

[0152] S1035. If the neighboring node is the node to be expanded and the actual improvement value corresponding to the neighboring node is greater than the actual improvement value corresponding to the current node to be processed, then the parent node corresponding to the neighboring node is determined as the current node to be processed.

[0153] In this embodiment of the application, this step is an optimization and adjustment step of the parent node of the node to be expanded. The core purpose is to ensure that the parent node of the node to be expanded is the current optimal choice, thereby ensuring that the actual cost of the path (improvement of actual cost value) is optimal.

[0154] Here, "neighborhood node to be expanded" means that the node has been included in the list of nodes to be expanded (Open list), but the optimal parent node has not yet been determined. By comparing the actual improvement value of the current neighboring node with the actual improvement value of the current node to be processed, if the actual improvement value of the neighboring node is greater, it means that using the current node to be processed as its parent node can obtain a better actual driving cost (shorter path length or lower turning cost). Therefore, the parent node of the neighboring node is updated to the current node to be processed, realizing dynamic optimization of the parent node and avoiding the actual cost of the path being too high due to improper initial parent node selection.

[0155] S1036. Repeatedly traverse the nodes to be processed until the target node is found.

[0156] In this embodiment of the application, this step is a cyclic execution link of node expansion. The core purpose is to continuously iterate the nodes to be processed and continuously update the nodes to be expanded by continuously repeating the process of the sub-steps S1032 to S1035 mentioned above, until the target node (path endpoint) is found, thus completing the core process of path search.

[0157] Among them, "repeated traversal" refers to the complete process of "selecting the current node to be processed, filtering neighboring nodes, processing neighboring nodes, and updating the node to be expanded" in a loop. In each loop, the node with the smallest target generation value is selected from the list of nodes to be expanded (Open list) as the new current node to be processed. The steps of filtering neighboring nodes, configuring nodes to be expanded, and optimizing parent nodes are repeated until the target node is found.

[0158] Furthermore, if the list of nodes to be expanded (Open list) is empty and the target node is still not found, the path planning is considered to have failed (the parameters need to be readjusted and the process retried).

[0159] S104. Starting from the target node, trace back the parent node level by level to determine the target driving path of the target vehicle in the environment.

[0160] Here, the target driving path refers to the complete driving path from the starting node to the target node that is planned by the method described in the embodiments of this application and meets the requirements of safety, feasibility, smoothness and efficiency.

[0161] In this step, once the target node is found, starting from the target node, all parent nodes are backtracked step by step in the order of "target node, parent node above, parent node above, parent node at each level, starting node". The backtracked nodes are arranged in order, collinear redundant nodes are removed, and the path is checked for collision-free and sharp turn-free, finally forming the target driving path of the target vehicle and outputting it to the vehicle control system.

[0162] In this embodiment of the application, this step converts the searched optimal node sequence into a complete driving path that the vehicle can execute, ensuring the integrity, feasibility and smoothness of the path, and providing a direct basis for the driving control of the vehicle.

[0163] In this way, the generated target driving path is complete, without breaks or redundant nodes, meeting the dynamic requirements of vehicle driving; the path is smooth and safe, avoiding high-risk obstacles, adapting to complex dynamic environments, and can be directly used for real-time driving control of vehicles, improving the reliability of autonomous driving.

[0164] For example, please refer to Figure 2 , Figure 2 This is an example diagram illustrating a vehicle driving path planning test provided in an embodiment of this application. Figure 2 As shown, the 40×40 grid map contains multiple black obstacles (walls / static obstacles) distributed in the upper and lower areas to simulate a complex urban / park road environment; the starting point of the path is the lower left corner area (coordinates approximately (2, 32)) and the ending point of the path is the upper right corner area (coordinates approximately (36, 10)); the red path is the path planned by the improved A* algorithm proposed in the embodiments of this application, and the blue path is the path planned by the traditional A* algorithm based only on the path length cost.

[0165] Among them, such as Figure 2As shown, in the upper half of the path (from the starting point to the middle obstacle zone), the red path makes only two smooth, large-angle turns when bypassing the middle obstacle, resulting in a smooth path; the blue path has more than four small inflection points, with obvious zigzag lines and more fragmented turning angles. In the lower half of the path (from the middle obstacle zone to the end point), the red path is next to a large obstacle on the right, with inflection points actively moving away from the obstacle edge to maintain a safe distance, resulting in smooth turns; the blue path is close to the obstacle edge, with more frequent inflection points, posing a risk of hitting the obstacle. The total length of the two paths is basically the same, indicating that the method described in this application achieves better smoothness and safety without increasing the path length.

[0166] The vehicle path planning method based on LiDAR and camera provided in this application obtains the semantic category and spatial location of obstacles through the fusion perception of LiDAR and camera. Based on the semantic category of the obstacle, the danger level of the obstacle is determined. The cost function of the improved A* algorithm introduces the dynamic heuristic cost corresponding to the danger level, the additional cost of turning angle, and the cost of potential field constraint. Then, the improved A* algorithm is used to determine the target driving path of the vehicle. It effectively solves the technical defects of the traditional A* algorithm, such as single perception information, poor path smoothness, weak adaptability to dynamic environment, and the inability to balance search efficiency and planning accuracy. It realizes safe, efficient and smooth path planning for vehicles in complex dynamic environments, significantly improves the reliability and adaptability of path planning, and fits various practical application scenarios of autonomous vehicles.

[0167] Please see Figure 3 , Figure 3 This is a schematic diagram of a vehicle path planning device based on LiDAR and a camera, provided as an embodiment of this application. Figure 3 As shown, the vehicle path planning device 300 includes: The semantic classification module 310 is used to collect perception data under a unified spatiotemporal reference corresponding to the environment where the target vehicle is located using the lidar and camera equipment set on the target vehicle, and to perform obstacle fusion perception and semantic classification on the perception data to determine the danger level of multiple obstacles in the environment where the target vehicle is located. The cost calculation module 320 is used to perform path planning based on the perceived data using a preset improved A* algorithm. During the node expansion process of the improved A* algorithm, it determines the dynamic heuristic cost value based on the dynamic parameters corresponding to the danger level and the distance between the nodes to be processed, and determines the improved actual cost value based on the path length value and turning angle value between the nodes to be processed. For the path inflection points in the nodes to be processed, it determines the potential field constraint cost value corresponding to the path inflection point based on the repulsive force exerted by the obstacles within the visible range on the path inflection point. Node search module 330 is used to traverse the nodes to be processed based on the improved actual value and the dynamic heuristic value corresponding to the node to be processed and the potential field constraint value corresponding to the path inflection point, so as to determine the parent node corresponding to each node to be processed and the target node in the nodes to be processed. The path generation module 340 is used to trace back from the target node to the parent node level by level to determine the target driving path of the target vehicle in the environment.

[0168] Furthermore, when the semantic classification module 310 is used to collect perception data under a unified spatiotemporal reference corresponding to the environment where the target vehicle is located using the lidar and camera equipment installed on the target vehicle, the semantic classification module 310 is used to: Spatiotemporal registration is performed on the lidar and camera equipment set on the target vehicle to establish the transformation relationship between the radar coordinate system corresponding to the lidar and the pixel coordinate system corresponding to the camera equipment. The raw perception data corresponding to the environment in which the target vehicle is located is collected using the lidar and camera equipment installed on the target vehicle. Based on the transformation relationship, the original sensing data is transformed to obtain sensing data under a unified spatiotemporal reference corresponding to the environment where the target vehicle is located.

[0169] Furthermore, when the semantic classification module 310 performs obstacle fusion perception and semantic classification on the perceived data to determine the danger levels corresponding to multiple obstacles in the environment where the target vehicle is located, the semantic classification module 310 is used to: Based on the radar perception data in the perception data, the lidar is used to determine the obstacle parameters corresponding to multiple obstacles in the environment where the target vehicle is located; The target detection model set in the camera device is used to identify multiple obstacles in the image perception data of the perception data, and the semantic category corresponding to each obstacle is determined; The obstacle parameters are fused and matched with the semantic category using multiple sensors to determine the location category information corresponding to each obstacle. Based on the location category information, the danger level of each obstacle is determined using a preset danger level classification rule.

[0170] Furthermore, when the cost calculation module 320 determines the dynamic heuristic cost value based on the dynamic parameters corresponding to the danger level and the distance between the nodes to be processed, the cost calculation module 320 is used to: In the nodes to be processed, a plurality of adjacent nodes to be processed are determined, and the node distance between each node to be processed and each of the adjacent nodes to be processed is determined. Based on the danger level corresponding to each of the adjacent nodes to be processed, determine the dynamic parameters corresponding to each of the adjacent nodes to be processed. Based on the distance between nodes and the dynamic parameters, the dynamic heuristic value corresponding to each node to be processed is determined.

[0171] Furthermore, when determining the actual improvement cost based on the path length and turning angle values ​​corresponding to the nodes to be processed, the cost calculation module 320 is used to: Determine the sum of the actual distances from each node to be processed to the starting node, and use the sum of the actual distances to determine the cumulative cost of the path length corresponding to each node to be processed; Based on the angle between the direction from the parent node to the node to be processed and the direction from the node to be processed to each of the adjacent nodes to be processed, determine the path turning angle value corresponding to each node to be processed; Based on the path turning angle value and the preset turning cost coefficient, determine the additional turning angle cost corresponding to each node to be processed; The sum of the cumulative cost of the path length and the additional cost of the turning angle is determined as the improved actual cost value corresponding to each node to be processed.

[0172] Furthermore, when the cost calculation module 320 determines the potential field constraint cost corresponding to a path inflection point based on the repulsive force exerted by obstacles within the visible range on the path inflection point in the node to be processed, the cost calculation module 320 is used to: For each path inflection point in the node to be processed, the straight-line Euclidean distance and repulsive force direction unit vector between each path inflection point and each obstacle within the visible range are determined, and the straight-line Euclidean distance is compared with a preset repulsive force distance threshold. When the straight-line Euclidean distance is less than or equal to the repulsive distance threshold, the repulsive force value of each obstacle in the visible range to each path inflection point is determined based on the straight-line Euclidean distance, the repulsive distance threshold, and the repulsive direction unit vector. When the straight-line Euclidean distance is greater than the repulsive distance threshold, the preset value is determined as the repulsive force value of each obstacle within the visible range on each path inflection point; Based on the repulsive force value, the total repulsive resultant force value corresponding to each path inflection point is determined, and the absolute value corresponding to the total repulsive resultant force value is determined as the potential field constraint cost value corresponding to the path inflection point.

[0173] Furthermore, when the node search module 330 is used to traverse the nodes to be processed based on the improved actual value and the dynamic heuristic value corresponding to the node to be processed, and the potential field constraint value corresponding to the path inflection point, in order to determine the parent node corresponding to each node to be processed and the target node among the nodes to be processed, the node search module 330 is used to: For a node to be processed that is not a path inflection point, the sum of the improved actual cost value and the dynamic heuristic cost value is determined as the target cost value corresponding to the node to be processed. For a node to be processed that is a path inflection point, the sum of the improved actual cost value, the dynamic heuristic cost value, and the potential field constraint cost value is determined as the target cost value corresponding to the node to be processed. Traverse the nodes to be processed, take the node to be processed corresponding to the first target value as the current node to be processed, and determine multiple neighboring nodes adjacent to the current node to be processed; wherein, when there are multiple nodes to be processed corresponding to the first target value, take the node to be processed corresponding to the first path turning angle value as the current node to be processed. For each of the neighboring nodes, if the neighboring node is an infeasible region node or has already been expanded, then no processing is performed on the neighboring node. If the neighboring node is not the node to be expanded, then the neighboring node is determined as the node to be expanded, and the current node to be processed is determined as the parent node corresponding to the neighboring node, and the target cost value corresponding to the neighboring node is determined. If the neighboring node is the node to be expanded and the actual improvement cost corresponding to the neighboring node is greater than the actual improvement cost corresponding to the current node to be processed, then the parent node corresponding to the neighboring node is determined as the current node to be processed. Repeatedly traverse the nodes to be processed until the target node is found.

[0174] The vehicle path planning device based on LiDAR and camera provided in this application obtains the semantic category and spatial location of obstacles through the fusion perception of LiDAR and camera. Based on the semantic category of the obstacle, the danger level of the obstacle is determined. The cost function of the improved A* algorithm introduces the dynamic heuristic cost corresponding to the danger level, the additional cost of turning angle, and the cost of potential field constraint. Then, the improved A* algorithm is used to determine the target driving path of the vehicle. It effectively solves the technical defects of the traditional A* algorithm, such as single perception information, poor path smoothness, weak adaptability to dynamic environment, and the inability to balance search efficiency and planning accuracy. It realizes safe, efficient and smooth path planning for vehicles in complex dynamic environments, significantly improves the reliability and adaptability of path planning, and fits various practical application scenarios of autonomous vehicles.

[0175] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device 400 includes a processor 410, a memory 420, and a bus 430.

[0176] The memory 420 stores machine-readable instructions executable by the processor 410. When the electronic device 400 is running, the processor 410 communicates with the memory 420 via the bus 430. When the machine-readable instructions are executed by the processor 410, they can perform the operations described above. Figure 1 The specific implementation of the vehicle path planning method based on LiDAR and camera in the method embodiment shown can be found in the method embodiment, and will not be repeated here.

[0177] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 The specific implementation of the vehicle path planning method based on LiDAR and camera in the method embodiment shown can be found in the method embodiment, and will not be repeated here.

[0178] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0179] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0180] 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 network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0181] In addition, 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.

[0182] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, 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 a portion 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, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0183] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A vehicle path planning method based on lidar and camera, characterized in that, The method includes: The LiDAR and camera equipment installed on the target vehicle are used to collect perception data under a unified spatiotemporal reference corresponding to the environment in which the target vehicle is located. The perception data is then subjected to obstacle fusion perception and semantic classification to determine the danger level of multiple obstacles in the environment in which the target vehicle is located. Based on the perceived data, a path planning is performed using a preset improved A* algorithm. During the node expansion process of the improved A* algorithm, the dynamic heuristic cost is determined based on the dynamic parameters corresponding to the danger level and the distance between the nodes to be processed. The improved actual cost is determined based on the path length and turning angle values ​​between the nodes to be processed. For the path inflection points in the nodes to be processed, the potential field constraint cost corresponding to the path inflection point is determined based on the repulsive force exerted by the obstacles within the visible range on the path inflection point. Based on the improved actual value and the dynamic heuristic value corresponding to the node to be processed, and the potential field constraint value corresponding to the path inflection point, the nodes to be processed are traversed to determine the parent node and the target node in each node to be processed. By tracing back from the target node to its parent node, the target driving path of the target vehicle in its environment is determined.

2. The method according to claim 1, characterized in that, The method of using lidar and camera equipment installed on the target vehicle to collect perception data under a unified spatiotemporal reference corresponding to the environment in which the target vehicle is located includes: Spatiotemporal registration is performed on the lidar and camera equipment set on the target vehicle to establish the transformation relationship between the radar coordinate system corresponding to the lidar and the pixel coordinate system corresponding to the camera equipment. The raw perception data corresponding to the environment in which the target vehicle is located is collected using the lidar and camera equipment installed on the target vehicle. Based on the transformation relationship, the original sensing data is transformed to obtain sensing data under a unified spatiotemporal reference corresponding to the environment where the target vehicle is located.

3. The method according to claim 1, characterized in that, The step of performing obstacle fusion perception and semantic classification on the perceived data to determine the danger levels corresponding to multiple obstacles in the environment where the target vehicle is located includes: Based on the radar perception data in the perception data, the lidar is used to determine the obstacle parameters corresponding to multiple obstacles in the environment where the target vehicle is located; The target detection model set in the camera device is used to identify multiple obstacles in the image perception data of the perception data, and the semantic category corresponding to each obstacle is determined; The obstacle parameters are fused and matched with the semantic category using multiple sensors to determine the location category information corresponding to each obstacle. Based on the location category information, the danger level of each obstacle is determined using a preset danger level classification rule.

4. The method according to claim 1, characterized in that, The determination of dynamic heuristic value based on the dynamic parameters corresponding to the danger level and the distance between nodes to be processed includes: In the nodes to be processed, a plurality of adjacent nodes to be processed are determined, and the node distance between each node to be processed and each of the adjacent nodes to be processed is determined. Based on the danger level corresponding to each of the adjacent nodes to be processed, determine the dynamic parameters corresponding to each of the adjacent nodes to be processed. Based on the distance between nodes and the dynamic parameters, the dynamic heuristic value corresponding to each node to be processed is determined.

5. The method according to claim 4, characterized in that, The determination of the actual cost of improvement based on the path length and turning angle values ​​between the nodes to be processed includes: Determine the sum of the actual distances from each node to be processed to the starting node, and use the sum of the actual distances to determine the cumulative cost of the path length corresponding to each node to be processed; Based on the angle between the direction from the parent node to the node to be processed and the direction from the node to be processed to each of the adjacent nodes to be processed, determine the path turning angle value corresponding to each node to be processed; Based on the path turning angle value and the preset turning cost coefficient, determine the additional turning angle cost corresponding to each node to be processed; The sum of the cumulative cost of the path length and the additional cost of the turning angle is determined as the improved actual cost value corresponding to each node to be processed.

6. The method according to claim 1, characterized in that, For path inflection points in the nodes to be processed, the potential field constraint cost corresponding to the path inflection point is determined based on the repulsive force exerted by obstacles within the visible range on the path inflection point, including: For each path inflection point in the node to be processed, the straight-line Euclidean distance and repulsive force direction unit vector between each path inflection point and each obstacle within the visible range are determined, and the straight-line Euclidean distance is compared with a preset repulsive force distance threshold. When the straight-line Euclidean distance is less than or equal to the repulsive distance threshold, the repulsive force value of each obstacle in the visible range to each path inflection point is determined based on the straight-line Euclidean distance, the repulsive distance threshold, and the repulsive direction unit vector. When the straight-line Euclidean distance is greater than the repulsive distance threshold, the preset value is determined as the repulsive force value of each obstacle within the visible range on each of the path inflection points; Based on the repulsive force value, the total repulsive resultant force value corresponding to each path inflection point is determined, and the absolute value corresponding to the total repulsive resultant force value is determined as the potential field constraint cost value corresponding to the path inflection point.

7. The method according to claim 1, characterized in that, The process of traversing the nodes to be processed based on the improved actual value, the dynamic heuristic value, and the potential constraint value corresponding to the path inflection point, to determine the parent node and the target node among the nodes to be processed for each node, includes: For a node to be processed that is not a path inflection point, the sum of the improved actual cost value and the dynamic heuristic cost value is determined as the target cost value corresponding to the node to be processed. For a node to be processed that is a path inflection point, the sum of the improved actual cost value, the dynamic heuristic cost value, and the potential field constraint cost value is determined as the target cost value corresponding to the node to be processed. Traverse the nodes to be processed, take the node to be processed corresponding to the first target value as the current node to be processed, and determine multiple neighboring nodes adjacent to the current node to be processed; wherein, when there are multiple nodes to be processed corresponding to the first target value, take the node to be processed corresponding to the first path turning angle value as the current node to be processed. For each of the neighboring nodes, if the neighboring node is an infeasible region node or has already been expanded, then no processing is performed on the neighboring node. If the neighboring node is not the node to be expanded, then the neighboring node is determined as the node to be expanded, and the current node to be processed is determined as the parent node corresponding to the neighboring node, and the target cost value corresponding to the neighboring node is determined. If the neighboring node is the node to be expanded and the actual improvement cost corresponding to the neighboring node is greater than the actual improvement cost corresponding to the current node to be processed, then the parent node corresponding to the neighboring node is determined as the current node to be processed. Repeatedly traverse the nodes to be processed until the target node is found.

8. A vehicle path planning device based on lidar and camera, characterized in that, The vehicle routing device includes: The semantic classification module is used to collect perception data under a unified spatiotemporal reference corresponding to the environment in which the target vehicle is located using the lidar and camera equipment installed on the target vehicle, and to perform obstacle fusion perception and semantic classification on the perception data to determine the danger level of multiple obstacles in the environment in which the target vehicle is located. The cost calculation module is used to perform path planning based on the perceived data using a preset improved A* algorithm. During the node expansion process of the improved A* algorithm, it determines the dynamic heuristic cost value based on the dynamic parameters corresponding to the danger level and the distance between the nodes to be processed, and determines the improved actual cost value based on the path length value and turning angle value between the nodes to be processed. For the path inflection points in the nodes to be processed, it determines the potential field constraint cost value corresponding to the path inflection point based on the repulsive force exerted by the obstacles within the visible range on the path inflection point. The node search module is used to traverse the nodes to be processed based on the improved actual value and the dynamic heuristic value corresponding to the node to be processed and the potential field constraint value corresponding to the path inflection point, so as to determine the parent node and the target node in the nodes to be processed for each node to be processed. The path generation module is used to trace back from the target node to the parent node level by level to determine the target driving path of the target vehicle in the environment.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. The machine-readable instructions are executed by the processor to perform the steps of the vehicle path planning method based on LiDAR and camera as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the vehicle path planning method based on lidar and camera as described in any one of claims 1 to 7.

Images