Mobile platform path planning and obstacle avoidance method and system based on negative obstacle perception

Abstract

The invention discloses a mobile platform path planning and obstacle avoidance method and system based on negative obstacle perception, and the method comprises the steps: collecting the image and depth data of an environment in front of the ground, recognizing a negative obstacle candidate region in a path, and extracting the geometric feature parameters of the negative obstacle candidate region; comparing the geometrical characteristic parameter of the negative obstacle with the platform structure parameter, and judging whether a preset crossing condition is met or not; when a crossing condition is met, the platform enters a crossing path mode, a plurality of candidate speed combinations are generated in a speed space, the candidate speed combinations are respectively substituted into a first path cost function for evaluation, and the candidate speed combination with the minimum cost is selected as control output; when the crossing condition is not met, the platform is switched to a bypassing path generation mode, the candidate paths are evaluated by using a second path cost function, and the path with the minimum cost is selected as an execution path; the method has the advantages that the requirement for negative obstacle navigation in complex and unstructured terrains is met, and the running safety and stability of the ground mobile platform are guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of mobile robot obstacle avoidance, and in particular to a mobile platform path planning and obstacle avoidance method and system based on negative obstacle perception. Background Art

[0002] The ability of ground mobile platforms to autonomously navigate complex environments relies on accurate perception of environmental obstacles and effective path planning. Traditional path planning systems typically employ a two-stage architecture of global path construction and local path correction, achieving relatively stable obstacle avoidance control in scenarios with static or regularized obstacles.

[0003] However, in real-world scenarios, ground depressions (i.e., negative obstacles), such as potholes, collapses, and trenches, often cannot be directly and effectively detected by conventional vision systems or lidar systems due to their low reflectivity, hidden shape, and difficulty appearing on elevation maps. Missing detection of negative obstacles often causes the platform to fall or become unstable, posing systemic operational risks. Therefore, how to accurately detect negative obstacles under real-time constraints and dynamically adjust the path accordingly has become a major technical challenge facing current intelligent navigation systems.

[0004] Existing studies have proposed methods for detecting negative obstacles based on image depth estimation, point cloud height difference analysis, or the fusion of depth maps and RGB images. Among them, some solutions use a RANSAC-based point cloud plane fitting method to model the ground, and determine points that deviate significantly from the fitting plane as obstacles or negative obstacle areas. For example, Chinese Patent Publication No. CN119860777A discloses a high-reliability positioning and navigation method for air-ground collaborative unmanned systems. Although such methods are relatively efficient in processing regular, flat roads, they have significant limitations when processing irregular terrain or naturally undulating roads. RANSAC-type methods are highly dependent on the flatness and continuity of the ground point cloud, and the fitting results are severely affected by noise points, slopes, and slight convexities. When the terrain has natural undulations, slope changes, or nonlinear distortions, the ground model fitting error increases significantly, which can easily lead to misclassification or blurred boundaries. In addition, such methods usually have difficulty in effectively extracting edge structures with broken geometric continuity and cannot accurately characterize obstacle boundaries. Moreover, since they process full-field point clouds, they have large computational complexity and high processing latency, making them unsuitable for high-frequency, real-time ground mobile platform application scenarios.

[0005] At the same time, existing negative obstacle recognition systems often focus only on point cloud feature extraction or image detection, lacking a collaborative coupling mechanism between image information and spatial point clouds. This fragmentation of the perception process limits the restoration of structural details and the improvement of boundary accuracy, and also makes it impossible for subsequent path planning to effectively rely on structured obstacle features for decision-making. Although some solutions propose to divide the path into patterns based on the proportional relationship between the width and depth of the negative obstacle and the size of the platform chassis, such as detouring, crossing, or slowing down to pass, these rules are mostly based on static threshold configurations, which make it difficult to make dynamic and reasonable adjustments based on the actual spatial form of the obstacle boundary, and have poor adaptability.

[0006] Therefore, there is currently a lack of an intelligent path planning method that integrates image-guided positioning, local point cloud modeling, and geometric structure analysis. Such a method should be able to extract accurate local spatial boundary information without relying on the assumption of global ground flatness. Combined with the platform's structural capabilities, it can achieve a linkage between traffic mode selection and dynamic path adjustment to meet the needs of negative obstacle navigation in complex, unstructured terrain, and ensure the safety and stability of ground mobile platform operations. Summary of the Invention

[0007] The technical problem to be solved by the present invention is how to provide a path planning and obstacle avoidance method that adapts to the negative obstacle navigation requirements in complex and unstructured terrain and ensures the safety and stability of the ground mobile platform operation.

[0008] The present invention solves the above technical problems through the following technical means: a mobile platform path planning and obstacle avoidance method based on negative obstacle perception, the method comprising:

[0009] Collect image and depth data of the ground environment ahead, identify candidate negative obstacle areas within the path, and extract their geometric feature parameters;

[0010] Comparing the geometric characteristic parameters of the negative obstacle with the platform structure parameters to determine whether the preset crossing conditions are met;

[0011] When the crossing condition is met, the platform enters the crossing path mode, generates several candidate speed combinations in the speed space, substitutes them into the first path cost function for evaluation, and selects the one with the lowest cost as the control output;

[0012] When the crossing condition is not met, the platform switches to the detour path generation mode, evaluates the candidate paths using the second path cost function, and selects the one with the lowest cost as the execution path;

[0013] When the maximum depth of the negative obstacle is less than the set pass depth threshold and lower than the ground clearance of the platform chassis, the platform enters the direct pass mode and passes through the current area directly without performing path adjustment.

[0014] The present invention compares the geometric characteristic parameters of the negative obstacle with the platform structural parameters to determine whether the crossing conditions are met. The first path cost function and the second path cost function are used to perform cost evaluation for the cases where the crossing conditions are met and not met, respectively, and the path with the smallest cost is selected as the execution path. The entire process switches between the crossing path mode, the detour path generation mode and the direct passing mode according to different situations, adapting to the negative obstacle navigation needs in complex and unstructured terrain, performing specific mode processing for different situations, and ensuring the safety and stability of the ground mobile platform operation.

[0015] Furthermore, the collecting of images and depth data of the ground front environment, identifying negative obstacle candidate areas within the path, and extracting their geometric feature parameters include:

[0016] Based on the environmental point cloud data collected by the ground-based lidar, a static environment map is generated using the SLAM module. Based on the known obstacle information in the map, the navigation starting point and target point are set, and a global path is generated using graph search or an extension method as a reference path for the platform execution. The platform executes the reference path and uses an RGB-D camera to detect the forward area image in real time. The image is input into the YOLOv8 network, and the bounding box information of the suspicious sunken area is output. The bounding box information is converted into camera space coordinates through the intrinsic and extrinsic parameter matrices, and then mapped to the three-dimensional point cloud data. The point cloud subset of the corresponding area is extracted, and the local mesh structure is generated using the triangular mesh reconstruction algorithm. The boundary outline of the negative obstacle is constructed, and its maximum horizontal span, boundary depth, and center position are extracted as geometric feature parameters.

[0017] Furthermore, the step of comparing the geometric characteristic parameters of the negative obstacle with the platform structure parameters to determine whether a preset crossing condition is met includes:

[0018] When the width of the negative obstacle is less than the wheelbase of the platform, the lateral offset between the center of the platform and the center of the obstacle is less than half the width of the obstacle, and the maximum depth of the obstacle is less than the ground clearance of the platform chassis, the preset crossing conditions are determined to be met.

[0019] Furthermore, the formula of the first path cost function is:

[0020] J(v,ω)=α·head(v,ω)+γ·vel(v,ω)+ε·crossover(v,ω)

[0021] Where v and ω represent the linear velocity and angular velocity, respectively; α, γ, and ε are weight coefficients; head(v, ω) represents the angle between the platform's current heading and the desired direction; vel(v, ω) represents the ratio of the current velocity to the maximum velocity; and crossover(v, ω) represents the lateral offset between the platform center and the obstacle center.

[0022] Furthermore, the formula for utilizing the second path cost function is:

[0023] G(v,ω)=α·head(v,ω)+β·dist(v,ω)+γ·vel(v,ω)+δ·Pdist(v,ω)

[0024] Where dist(v,ω) represents the straight-line distance between the current position and the target point, Pdist(v,ω) represents the average deviation between the detour path and the original path; β and δ are both weight factors.

[0025] Furthermore, the method further comprises:

[0026] Based on the cost evaluation and path selection results, the local path adjustment algorithm is combined to generate the speed control pair of the platform at the current moment and send it to the underlying driver to control the platform movement and control the platform execution path in real time. Among them, the local path adjustment algorithm is the dynamic window method, and the speed control pair is the linear velocity and angular velocity.

[0027] Furthermore, the method further comprises:

[0028] The platform has a built-in state machine mechanism to monitor the deviation between the platform's current position and the expected path point, and the changes in the distribution of the environmental cost map in real time, and to switch the platform's mode.

[0029] The present invention also provides a mobile platform path planning and obstacle avoidance system based on negative obstacle perception, the system comprising:

[0030] The perception module is used to collect image and depth data of the ground environment ahead, identify candidate negative obstacle areas within the path, and extract their geometric feature parameters;

[0031] A geometric analysis module is used to compare the geometric characteristic parameters of the negative obstacle with the platform structure parameters to determine whether the preset crossing conditions are met;

[0032] The path evaluation module is used to enable the platform to enter the crossing path mode when the crossing conditions are met, generate several candidate speed combinations in the speed space, substitute them into the first path cost function for evaluation, and select the one with the lowest cost as the control output; when the crossing conditions are not met, the platform switches to the detour path generation mode, evaluates the candidate paths using the second path cost function, and selects the one with the lowest cost as the execution path; when the maximum depth of the negative obstacle is less than the set pass depth threshold and lower than the ground clearance of the platform chassis, the platform enters the direct passing mode and directly passes through the current area without performing path adjustment.

[0033] Furthermore, the sensing module is also used to:

[0034] Based on the environmental point cloud data collected by the ground-based lidar, a static environment map is generated using the SLAM module. Based on the known obstacle information in the map, the navigation starting point and target point are set, and a global path is generated using graph search or an extension method as a reference path for the platform execution. The platform executes the reference path and uses an RGB-D camera to detect the forward area image in real time. The image is input into the YOLOv8 network, and the bounding box information of the suspicious sunken area is output. The bounding box information is converted into camera space coordinates through the intrinsic and extrinsic parameter matrices, and then mapped to the three-dimensional point cloud data. The point cloud subset of the corresponding area is extracted, and the local mesh structure is generated using the triangular mesh reconstruction algorithm. The boundary outline of the negative obstacle is constructed, and its maximum horizontal span, boundary depth, and center position are extracted as geometric feature parameters.

[0035] Furthermore, the geometric analysis module is also used to:

[0036] When the width of the negative obstacle is less than the wheelbase of the platform, the lateral offset between the center of the platform and the center of the obstacle is less than half the width of the obstacle, and the maximum depth of the obstacle is less than the ground clearance of the platform chassis, the preset crossing conditions are determined to be met.

[0037] Furthermore, the formula of the first path cost function is:

[0038] J(v,ω)=α·head(v,ω)+γ·vel(v,ω)+ε·crossover(v,ω)

[0039] Where v and ω represent the linear velocity and angular velocity, respectively; α, γ, and ε are weight coefficients; head(v, ω) represents the angle between the platform's current heading and the desired direction; vel(v, ω) represents the ratio of the current velocity to the maximum velocity; and crossover(v, ω) represents the lateral offset between the platform center and the obstacle center.

[0040] Furthermore, the formula for utilizing the second path cost function is:

[0041] G(v,ω)=α·head(v,ω)+β·dist(v,ω)+γ·vel(v,ω)+δ·Pdist(v,ω)

[0042] Where dist(v,ω) represents the straight-line distance between the current position and the target point, Pdist(v,ω) represents the average deviation between the detour path and the original path; β and δ are both weight factors.

[0043] Furthermore, the system further includes a path control module, which is configured to:

[0044] Based on the cost evaluation and path selection results, the local path adjustment algorithm is combined to generate the speed control pair of the platform at the current moment and send it to the underlying driver to control the platform movement and control the platform execution path in real time. Among them, the local path adjustment algorithm is the dynamic window method, and the speed control pair is the linear velocity and angular velocity.

[0045] Furthermore, the system further includes a state control module, which is further configured to:

[0046] The platform has a built-in state machine mechanism to monitor the deviation between the platform's current position and the expected path point, and the changes in the distribution of the environmental cost map in real time, and to switch the platform's mode.

[0047] The advantages of the present invention are:

[0048] (1) The present invention compares the geometric characteristic parameters of the negative obstacle with the platform structural parameters to determine whether the crossing conditions are met. The first path cost function and the second path cost function are used to evaluate the cost when the crossing conditions are met and when the crossing conditions are not met, respectively, and the path with the smallest cost is selected as the execution path. The entire process switches between the crossing path mode, the detour path generation mode and the direct passing mode according to different situations, adapting to the negative obstacle navigation needs in complex and unstructured terrain, performing specific mode processing for different situations, and ensuring the safety and stability of the ground mobile platform operation.

[0049] (2) The present invention combines image perception, point cloud structure analysis and kinematic rule judgment, which not only improves the detection accuracy and boundary restoration capability of negative obstacles, but also takes into account path safety and actual execution efficiency. It is suitable for autonomous navigation systems of ground mobile robots, self-driving vehicles and other intelligent platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of the mobile platform path planning and obstacle avoidance method based on negative obstacle perception disclosed in Example 1 of the present invention;

[0051] Figure 2This is the entire process of path decision-making and control execution of the mobile platform path planning and obstacle avoidance method based on negative obstacle perception disclosed in Example 1 of the present invention after detecting a negative obstacle, wherein: Figure 2 (a) is the overall picture of the experimental scene. Figure 2 (b) is the negative obstacle candidate area identified by the visual detection module; Figure 2 (c) is the residual cloud extraction result. Figure 2 (d) is the pit boundary model generated by the contour reconstruction algorithm. Figure 2 (e) The mobile platform is at its initial position. Figure 2 (f) is the target position, Figure 2 (g) The platform drives along the original path to the pit area and gets stuck without triggering the obstacle avoidance strategy; Figure 2 (h) The process of the platform successfully bypassing the pit and rejoining the global path after triggering the dynamic obstacle avoidance strategy;

[0052] Figure 3 This is the path adjustment strategy adopted by the mobile platform path planning and obstacle avoidance method based on negative obstacle perception disclosed in Example 1 of the present invention when facing different types of negative obstacles, wherein: Figure 3 (a) is the pit detection and positioning diagram, Figure 3 (b) is the residual cloud extraction result. Figure 3 (c) is the contour reconstruction result. Figure 3 (d) shows the trajectory of the platform successfully crossing the pit under the guidance of the dynamic window control strategy, with an enlarged image showing the spatial relationship between the platform and the obstacle edge. Figure 3 (e) is the detection area map, Figure 3 (f) is the point cloud reconstruction result, Figure 3 (g) is the overall view of the experimental environment. Figure 3 (h) The platform automatically plans a detour path and completes the obstacle avoidance process after detecting that the pit is off-center. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0054] Example 1

[0055] Embodiment 1 of the present invention provides a mobile platform path planning and obstacle avoidance method based on negative obstacle perception, which can be deployed on a ground mobile platform with non-holonomic motion constraints, such as a robot or vehicle platform with an Ackerman steering structure. The platform is equipped with devices such as a laser radar, a depth perception camera, and an embedded control unit, and can complete tasks such as autonomous positioning, map construction, path planning, environmental perception, and obstacle avoidance control. In order to verify the effectiveness of the solution of the present invention, a test environment containing typical negative obstacles (such as pits and sunken areas) can be constructed in a real field for application deployment. Figure 1 The overall processing flow of the path planning system of the present invention is illustrated, which includes three stages: the first stage is the global path planning stage, in which the system performs path generation and path smoothing based on the constructed static map to obtain a preliminary feasible path; the second stage is the negative obstacle perception stage, in which the system identifies potential pit areas ahead through visual detection and point cloud analysis, completes the extraction of interest regions, geometric dimension calculation and boundary completion; the third stage is the intelligent path decision stage, in which the system automatically determines whether the crossing conditions are met based on the width and position offset of the negative obstacle, and dynamically adjusts the path based on different strategies. Figure 1 , the method comprises the following steps:

[0056] S1. During the system initialization phase, the platform scans the surrounding scene through the configured environmental perception device to generate an environmental model containing information about static obstacles above the ground. This model can be constructed into a two-dimensional grid map through a certain synchronous positioning and mapping method, and its resolution can be set to the centimeter level to meet the accuracy requirements of path search and local obstacle avoidance. For example, based on the environmental point cloud data collected by ground-based laser radar (such as 2D / 3D LiDAR), a static environmental map is generated in combination with the synchronous positioning and mapping algorithm (SLAM, Simultaneous Localization and Mapping). The SLAM module can use existing algorithms such as GMapping, Cartographer, LOAM, etc. to achieve platform pose tracking through feature extraction, matching and motion estimation, and use the pose relationship between laser frames for global map splicing. This map usually only contains information about positive obstacles above the ground and does not cover sunken areas on the ground.

[0057] S2. The system receives the start and destination points and generates a reference path from the start to the destination using either a graph-based search strategy (e.g., A*, Dijkstra) or a sampling-based path search strategy (e.g., RRT, PRM). This generated path is collision-free and relies primarily on positive obstacle information in the map. It does not include the identification and handling of subsidence structures, but primarily ensures spatial accessibility.

[0058] S3. When the platform is performing tasks along the reference path, the system runs the perception module in real time to detect whether there are negative obstacles ahead of the navigation path. The perception module may include two submodules: image perception and spatial data fusion. The image perception module is used to identify sunken areas on the ground, and its processing flow may include steps such as image preprocessing, region segmentation, and discrimination classification. After detecting the candidate negative obstacle area, the system further calls the spatial perception module, projects the corresponding image position into the depth map or point cloud data, and extracts its three-dimensional spatial structure. The spatial structure can be obtained by methods such as ground point separation and boundary reconstruction, and finally the maximum horizontal width h of the negative obstacle is extracted. ob , maximum depth h, boundary shape and geometric center position (x c ,y c ) and other parameters. In this embodiment, the above-mentioned three-dimensional spatial structure of the negative obstacle can be achieved by the following refined method. Specifically, during the path tracking process, the forward area of ​​the platform is detected in real time by the visual perception module. The module includes an RGB-D camera (such as Intel RealSense D435i) installed at the front end of the platform, which cooperates with the target detection network based on deep learning to quickly identify possible ground depressions. The network can adopt YOLOv8 or a similar lightweight structure, with the input being an RGB image with a ground scene and a corresponding depth map, and the output being the bounding box information (bounding box) of the suspicious depression area in the two-dimensional image. The network training adopts a negative obstacle sample image annotation dataset, and the target category is a negative obstacle type with obvious visual features such as "potholes".

[0059] S4. The system converts the target area in the above-mentioned image coordinate system into camera space coordinates through the intrinsic and extrinsic matrix, and then maps it to the three-dimensional point cloud data to extract the point cloud subset of the corresponding area. Subsequently, a triangular mesh reconstruction algorithm (such as the Delaunay method) is used in this local point cloud to generate a local mesh structure and calculate the normal vector of each face in the mesh. By comparing the angle or change rate between the normal vector of a certain triangle face and the normal vector of its adjacent face, the normal mutation area is determined, and the edge points are further extracted, and the boundary contour of the negative obstacle is constructed accordingly.

[0060] S5, the system is based on the geometric characteristics of the obstacle and the structural parameters of the platform itself (such as wheelbase W, ground clearance h o During the decision-making process, the platform may have three operating modes: crossing mode, detour mode, and direct pass mode. The mode determination mechanism is as follows:

[0061] When the width of the negative obstacle is less than the platform's wheelbase, the lateral offset between the platform center and the obstacle center is less than half the obstacle width, and the maximum depth of the obstacle is less than the platform chassis ground clearance, the system determines that the area is crossable. At this time, the system will enter the cross-path mode. To ensure that the platform has good posture stability and path smoothness when passing through the negative obstacle area, the system defines the cost function of the cross-path as follows:

[0062] J(v,ω)=α·head(v,ω)+γ·vel(v,ω)+ε·crossover(v,ω)

[0063] Among them, v and ω represent linear velocity and angular velocity respectively, and α, γ, and ε are weight coefficients used to balance the contribution of various indicators. The specific meanings of the indicators are as follows:

[0064] head(v,ω): represents the angle between the platform's current heading and the desired direction, which can be defined as |θ t -θ g |, where θ t is the current heading angle of the platform, θ g is the global path direction;

[0065] vel(v,ω): represents the ratio of current speed to maximum speed, usually in normalized form;

[0066] crossover(v,ω): represents the lateral offset between the platform center and the obstacle center, which is used to reflect whether the platform is aligned with the center area of ​​the pit.

[0067] The system generates several candidate speed combinations in the speed space, substitutes them into the above cost function for evaluation, and selects the one with the smallest cost as the control output, thereby guiding the platform to pass through the negative obstacle area stably.

[0068] If the size of the detected negative obstacle is larger than the platform structure's allowed crossing limit, or the obstacle has a large lateral offset, or its depth significantly exceeds the chassis ground clearance, that is, if the above-mentioned conditions for crossing are not met, the system will determine that the obstacle is not crossable and automatically switch to detour path generation mode. In this mode, the system needs to temporarily generate an alternative path outside the obstacle area. To improve the compactness of the detour path and the regression ability of the final path, the system constructs the detour path evaluation function as follows:

[0069] G(v,ω)=α·head(v,ω)+β·dist(v,ω)+γ·vel(v,ω)+δ·Pdist(v,ω)

[0070] Where dist(v, ω) represents the straight-line distance between the current location and the target point, and Pdist(v, ω) represents the average deviation between the detour path and the original path. β and δ are weighting factors that control the trade-off between path deviation and path length. The path evaluation process is similar to that of the leapfrogging mode. By comparing candidate paths, the path with the lowest cost is selected as the execution path.

[0071] For areas with shallow depressions and small widths, the system can set a pass depth threshold h thresh When the maximum obstacle depth is less than this threshold and within the dual constraints of wheelbase and ground clearance, the system considers the area traversable and does not perform path adjustments. This strategy can avoid redundant path adjustments and improve mission execution efficiency, but in actual application, it should be comprehensively evaluated in combination with platform load stability and mission requirements.

[0072] S6. Based on the above path evaluation and selection results, the system combines the local path adjustment algorithm (such as the dynamic window method DWA) to generate the speed control pair (linear velocity and angular velocity) of the platform at the current moment, and controls the platform execution path in real time.

[0073] S7. Path strategy judgment and control mode switching are implemented by the state machine module. The state machine mechanism monitors the deviation between the platform's current position and the expected path point, as well as changes in the distribution of the environmental cost map, in real time. It switches between "path tracking" mode and "obstacle avoidance adjustment" mode based on the current state, ensuring that the platform automatically corrects its motion behavior in the presence of negative obstacles or path deviations. The state machine includes path tracking state, path adjustment state, and path regression state. During each control cycle, the system evaluates the environmental state of the area ahead of the path and, based on this, determines whether to switch to the obstacle avoidance path generation process. Once the obstacle is removed, the system guides the platform to gradually return to the original path, achieving closed-loop control and path reconnection.

[0074] Figure 2 The whole process of path decision and control execution of the system after detecting negative obstacles is demonstrated. Figure 2 (a) is the overall picture of the experimental scene. Figure 2 (b) is the negative obstacle candidate area identified by the visual detection module; Figure 2 (c) is the residual cloud extraction result. Figure 2 (d) is the pit boundary model generated by the contour reconstruction algorithm. Figure 2 (e) The mobile platform is at its initial position. Figure 2 (f) is the target position, Figure 2 (g) The platform drives along the original path to the pit area and gets stuck without triggering the obstacle avoidance strategy; Figure 2 (h) The process of the platform successfully bypassing the pit and rejoining the global path after triggering the dynamic obstacle avoidance strategy.

[0075] Figure 3 The comparison shows the path adjustment strategies adopted by the system when facing different types of negative obstacles. Figure 3 (a–d) are the experimental processes of the spanning strategy, where Figure 3 (a) is the pit detection and positioning diagram, Figure 3 (b) is the residual cloud extraction result. Figure 3 (c) is the contour reconstruction result. Figure 3 (d) The motion trajectory of the platform successfully crossing the pit under the guidance of the dynamic window control strategy, with an enlarged image showing the spatial relationship between the platform and the obstacle edge. Figure 3 (e–h) are the experimental processes of the detour strategy, where Figure 3 (e) is the detection area map, Figure 3 (f) is the point cloud reconstruction result, Figure 3 (g) is the overall view of the experimental environment. Figure 3 (h) shows the platform automatically planning a detour path and completing the obstacle avoidance process after detecting that the pit is off-center. The figure superimposes the visual detection frame recognition results. Figure 3 and Figure 2 Comparative analysis shows that the method of the present invention realizes the detection and boundary restoration of negative obstacles, as well as point cloud reconstruction, path planning, obstacle avoidance processing, etc. in corresponding scenarios. The entire method integrates image-guided positioning, point cloud local modeling and geometric structure analysis.

[0076] Through the above technical solution, the present invention provides a path decision-making method suitable for environments with ground depressions. Combining image perception, point cloud structure analysis and kinematic rule judgment, it not only improves the detection accuracy and boundary restoration capability of negative obstacles, but also takes into account path safety and actual execution efficiency. It is suitable for autonomous navigation systems of ground mobile robots, self-driving vehicles and other intelligent platforms.

[0077] Example 2

[0078] Based on Example 1, Example 2 of the present invention further provides a mobile platform path planning and obstacle avoidance system based on negative obstacle perception, the system comprising:

[0079] The perception module is used to collect image and depth data of the ground environment ahead, identify candidate negative obstacle areas within the path, and extract their geometric feature parameters;

[0080] A geometric analysis module is used to compare the geometric characteristic parameters of the negative obstacle with the platform structure parameters to determine whether the preset crossing conditions are met;

[0081] The path evaluation module is used to enable the platform to enter the crossing path mode when the crossing conditions are met, generate several candidate speed combinations in the speed space, substitute them into the first path cost function for evaluation, and select the one with the lowest cost as the control output; when the crossing conditions are not met, the platform switches to the detour path generation mode, evaluates the candidate paths using the second path cost function, and selects the one with the lowest cost as the execution path; when the maximum depth of the negative obstacle is less than the set pass depth threshold and lower than the ground clearance of the platform chassis, the platform enters the direct passing mode and directly passes through the current area without performing path adjustment.

[0082] Specifically, the perception module is also used to:

[0083] Based on the environmental point cloud data collected by the ground-based lidar, a static environment map is generated using the SLAM module. Based on the known obstacle information in the map, the navigation starting point and target point are set, and a global path is generated using graph search or an extension method as a reference path for the platform execution. The platform executes the reference path and uses an RGB-D camera to detect the forward area image in real time. The image is input into the YOLOv8 network, and the bounding box information of the suspicious sunken area is output. The bounding box information is converted into camera space coordinates through the intrinsic and extrinsic parameter matrices, and then mapped to the three-dimensional point cloud data. The point cloud subset of the corresponding area is extracted, and the local mesh structure is generated using the triangular mesh reconstruction algorithm. The boundary outline of the negative obstacle is constructed, and its maximum horizontal span, boundary depth, and center position are extracted as geometric feature parameters.

[0084] Specifically, the geometric analysis module is further used to:

[0085] When the width of the negative obstacle is less than the wheelbase of the platform, the lateral offset between the center of the platform and the center of the obstacle is less than half the width of the obstacle, and the maximum depth of the obstacle is less than the ground clearance of the platform chassis, the preset crossing conditions are determined to be met.

[0086] Specifically, the formula of the first path cost function is:

[0087] J(v,ω)=α·head(v,ω)+γ·vel(v,ω)+ε·crossover(v,ω)

[0088] Where v and ω represent the linear velocity and angular velocity, respectively; α, γ, and ε are weight coefficients; head(v, ω) represents the angle between the platform's current heading and the desired direction; vel(v, ω) represents the ratio of the current velocity to the maximum velocity; and crossover(v, ω) represents the lateral offset between the platform center and the obstacle center.

[0089] More specifically, the formula for utilizing the second path cost function is:

[0090] G(v,ω)=α·head(v,ω)+β·dist(v,ω)+γ·vel(v,ω)+δ·Pdist(v,ω)

[0091] Where dist(v,ω) represents the straight-line distance between the current position and the target point, Pdist(v,ω) represents the average deviation between the detour path and the original path; β and δ are both weight factors.

[0092] Specifically, the system further includes a path control module, which is used to:

[0093] Based on the cost evaluation and path selection results, the local path adjustment algorithm is combined to generate the speed control pair of the platform at the current moment and send it to the underlying driver to control the platform movement and control the platform execution path in real time. Among them, the local path adjustment algorithm is the dynamic window method, and the speed control pair is the linear velocity and angular velocity.

[0094] Specifically, the system further includes a state control module, which is further configured to:

[0095] The platform has a built-in state machine mechanism to monitor the deviation between the platform's current position and the expected path point, and the changes in the distribution of the environmental cost map in real time, and to switch the platform's mode.

[0096] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A mobile platform path planning and obstacle avoidance method based on negative obstacle perception, characterized in that: The method comprises: Collect image and depth data of the ground environment ahead, identify candidate negative obstacle areas within the path, and extract their geometric feature parameters; Comparing the geometric characteristic parameters of the negative obstacle with the platform structure parameters to determine whether the preset crossing conditions are met; When the crossing condition is met, the platform enters the crossing path mode, generates several candidate speed combinations in the speed space, substitutes them into the first path cost function for evaluation, and selects the one with the lowest cost as the control output; When the crossing condition is not met, the platform switches to the detour path generation mode, evaluates the candidate paths using the second path cost function, and selects the one with the lowest cost as the execution path; When the maximum depth of the negative obstacle is less than the set pass depth threshold and lower than the ground clearance of the platform chassis, the platform enters the direct pass mode and passes through the current area directly without performing path adjustment.

2. The mobile platform path planning and obstacle avoidance method based on negative obstacle perception according to claim 1, characterized in that: The process of collecting image and depth data of the front ground environment, identifying candidate negative obstacle areas within the path, and extracting their geometric feature parameters includes: Based on the environmental point cloud data collected by the ground-based lidar, a static environment map is generated using the SLAM module. Based on the known obstacle information in the map, the navigation starting point and target point are set, and a global path is generated using graph search or an extension method as a reference path for the platform execution. The platform executes the reference path and uses an RGB-D camera to detect the forward area image in real time. The image is input into the YOLOv8 network, and the bounding box information of the suspicious sunken area is output. The bounding box information is converted into camera space coordinates through the intrinsic and extrinsic parameter matrices, and then mapped to the three-dimensional point cloud data. The point cloud subset of the corresponding area is extracted, and the local mesh structure is generated using the triangular mesh reconstruction algorithm. The boundary outline of the negative obstacle is constructed, and its maximum horizontal span, boundary depth, and center position are extracted as geometric feature parameters.

3. The mobile platform path planning and obstacle avoidance method based on negative obstacle perception according to claim 1, characterized in that: The step of comparing the geometric characteristic parameters of the negative obstacle with the platform structure parameters to determine whether a preset crossing condition is met includes: When the width of the negative obstacle is less than the wheelbase of the platform, the lateral offset between the center of the platform and the center of the obstacle is less than half the width of the obstacle, and the maximum depth of the obstacle is less than the ground clearance of the platform chassis, the preset crossing conditions are determined to be met.

4. The mobile platform path planning and obstacle avoidance method based on negative obstacle perception according to claim 1, characterized in that: The formula of the first path cost function is: J(v,ω)=α·head(v,ω)+γ·vel(v,ω)+ε·crossover(v,ω) Where v and ω represent the linear velocity and angular velocity, respectively; α, γ, and ε are weight coefficients; head(v, ω) represents the angle between the platform's current heading and the desired direction; vel(v, ω) represents the ratio of the current velocity to the maximum velocity; and crossover(v, ω) represents the lateral offset between the platform center and the obstacle center.

5. The mobile platform path planning and obstacle avoidance method based on negative obstacle perception according to claim 4 is characterized in that: The formula for utilizing the second path cost function is: G(v,ω)=α·head(v,ω)+β·dist(v,ω)+γ·vel(v,ω)+δ·Pdist(v,ω) Where dist(v,ω) represents the straight-line distance between the current position and the target point, Pdist(v,ω) represents the average deviation between the detour path and the original path; β and δ are both weight factors.

6. The mobile platform path planning and obstacle avoidance method based on negative obstacle perception according to claim 1, characterized in that: The method further comprises: Based on the cost evaluation and path selection results, the local path adjustment algorithm is combined to generate the speed control pair of the platform at the current moment and send it to the underlying driver to control the platform movement and control the platform execution path in real time. Among them, the local path adjustment algorithm is the dynamic window method, and the speed control pair is the linear velocity and angular velocity.

7. The mobile platform path planning and obstacle avoidance method based on negative obstacle perception according to claim 1, characterized in that: The method further comprises: The platform has a built-in state machine mechanism to monitor the deviation between the platform's current position and the expected path point, and the changes in the distribution of the environmental cost map in real time, and to switch the platform's mode.

8. Mobile platform path planning and obstacle avoidance system based on negative obstacle perception, characterized by: The system comprises: The perception module is used to collect image and depth data of the ground environment ahead, identify candidate negative obstacle areas within the path, and extract their geometric feature parameters; A geometric analysis module is used to compare the geometric characteristic parameters of the negative obstacle with the platform structure parameters to determine whether the preset crossing conditions are met; The path evaluation module is used to enable the platform to enter the crossing path mode when the crossing conditions are met, generate several candidate speed combinations in the speed space, substitute them into the first path cost function for evaluation, and select the one with the lowest cost as the control output; when the crossing conditions are not met, the platform switches to the detour path generation mode, evaluates the candidate paths using the second path cost function, and selects the one with the lowest cost as the execution path; when the maximum depth of the negative obstacle is less than the set pass depth threshold and lower than the ground clearance of the platform chassis, the platform enters the direct passing mode and directly passes through the current area without performing path adjustment.

9. The mobile platform path planning and obstacle avoidance system based on negative obstacle perception according to claim 8, characterized in that: The perception module is also used to: Based on the environmental point cloud data collected by the ground-based lidar, the SLAM module is used to generate a static environmental map. Based on the known obstacle information in the map, the navigation starting point and target point are set, and a global path is generated using graph search or expansion methods as a reference path for the platform execution. The platform executes the reference path and uses an RGB-D camera to detect the forward area image in real time. The image is input into the YOLOv8 network, and the bounding box information of the suspected sunken area is output. The bounding box information is converted into camera space coordinates through the intrinsic and extrinsic parameter matrices, and then mapped to the three-dimensional point cloud data. The point cloud subset of the corresponding area is extracted, and the local mesh structure is generated using the triangular mesh reconstruction algorithm. The boundary outline of the negative obstacle is constructed, and its maximum horizontal span, boundary depth, and center position are extracted as geometric feature parameters.

10. The mobile platform path planning and obstacle avoidance system based on negative obstacle perception according to claim 8, characterized in that: The geometric analysis module is also used to: When the width of the negative obstacle is less than the wheelbase of the platform, the lateral offset between the center of the platform and the center of the obstacle is less than half the width of the obstacle, and the maximum depth of the obstacle is less than the ground clearance of the platform chassis, the preset crossing conditions are determined to be met.

Images