Path planning method for automatic driving sweeping vehicle in complex narrow road scene

Abstract

The invention discloses a narrow road passing path planning method suitable for an automatic driving sweeping vehicle. The narrow road passing path planning method is particularly suitable for a complex unstructured environment in an open space. The method is based on dynamic programming mapping expansion, and comprises the following steps: firstly, constructing a grid map containing obstacles, generating a global cost graph by using breadth-first search, and constructing a Voronoi graph through obstacle distance transformation; on this basis, a heuristic path search algorithm is used to generate a preliminary path, and a potential narrow lane area in the path is extracted. And then, kinematics compensation is performed on the path by adopting an expansion mechanism based on a bicycle model, the expansion resolution is improved in a narrow lane area, and path search is guided in combination with a composite cost function. And finally, an optimal path meeting the requirements of safety, smoothness and traffic efficiency is obtained. The method improves the path planning robustness and real-time performance in a narrow channel environment, and is suitable for an automatic driving system in a complex urban environment and a semi-closed space.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automatic driving, and particularly relates to a path search method based on dynamic programming mapping expansion, which is suitable for efficient passing path planning tasks of an automatic driving sweeper in a complex narrow lane scene. BACKGROUND

[0002] Currently, sweeping automatic driving vehicles are widely used in open but complex structured environments such as parks, footpaths and underground parking lots. However, there are often problems such as non-linear obstacle boundaries, unpredictable object placement and narrow passing gaps in such environments, which seriously restrict the performance of traditional path planning algorithms.

[0003] Although the traditional method based on reference line optimization can generate continuous and smooth trajectories, it often fails to pass in unstructured scenes due to unavailable paths or attitude constraint conflicts. Although the existing search-based algorithm (such as A* and D*) has advantages in diversity, the path search efficiency and trajectory quality are difficult to balance in narrow lane environments, especially when there are turning restrictions and vehicle dynamic constraints, which may cause problems such as "search not converging" or "trajectory unable to execute".

[0004] In addition, most current path search systems rely on upstream perception to provide structured environment labeling, and lack the ability to directly identify narrow lanes from sparse images or grid maps and quickly adapt, resulting in poor generalization ability in complex scenes.

[0005] Therefore, there is an urgent need for a path search method with scene understanding ability, dynamic resolution adjustment ability and compatibility with kinematic constraints to enhance the passing ability and system stability in complex and narrow environments.

[0006] Therefore, the prior art still needs to be improved. SUMMARY

[0007] In view of the above deficiencies of the prior art, the purpose of the present application is to provide a path search method based on dynamic programming mapping expansion, which aims to solve the following technical problems:

[0008] 1. Trajectories cannot cover the passing needs of narrow scenes;

[0009] 2. search time and accuracy are difficult to balance;

[0010] 3. lack of robust control redundancy for vehicle attitude.

[0011] The technical solution of the present application is as follows:

[0012] The path planning method of the automatic driving sweeper vehicle in a complex narrow lane scene provided by the present application comprises:

[0013] S1, construct a two-dimensional grid map containing obstacles, and set the grid resolution;

[0014] S2, generate a global cost map guided by the end point based on breadth-first search;

[0015] S3, construct a Viterbi graph by obstacle distance transform;

[0016] S4, generate a preliminary path using A* algorithm;

[0017] S5, extract a narrow region that meets the preset condition based on the path;

[0018] S6, expand the path in the narrow region using an expansion model, which is a bicycle model considering the front wheel steering angle;

[0019] S7, calculate the total cost of the path according to a composite cost function, wherein the total cost includes path distance cost, gear shift penalty, steering shift penalty, and obstacle distance cost;

[0020] S8, dynamically improve the path expansion resolution in the narrow region, and output the optimal path.

[0021] In one embodiment, the construction of the grid map discretizes the continuous space into a two-dimensional matrix, and the obstacle region is marked as 1 and the passable region is marked as 0.

[0022] In one embodiment, the global cost map calculates the cost value of each grid to the target end point based on L1 norm, which is used to guide the path search direction.

[0023] In one embodiment, the Viterbi graph is calculated based on the Euclidean distance of each grid to the obstacle, and the distance is used for the cost design of obstacle avoidance.

[0024] In one embodiment, the scoring function of the A* algorithm is:

[0025] f(n) = C_{\text{L1}}(n) + \alpha \cdot D_{\text{obs}}(n);

[0026] wherein C_{\text{L1}} is the end point cost, D_{\text{obs}} is the obstacle distance, and \alpha is the weight coefficient.

[0027] In one embodiment, the extraction of the narrow region is based on the neighborhood sparsity of the grid in the path, and if the number of passable grids in the neighborhood is lower than the preset threshold, it is determined as a narrow passage.

[0028] In one embodiment, the extended model includes lateral position, longitudinal position, heading angle, speed and front wheel steering angle of the vehicle, and motion prediction is made based on discrete time steps.

[0029] In one embodiment, the total cost function f(n) includes:

[0030] Euclidean distance cost with parent node;

[0031] Gear shift penalty term;

[0032] Front wheel steering angle switching from 0 to non-zero penalty term;

[0033] Front wheel left-right switching penalty term.

[0034] In one embodiment, the heuristic cost h(n) includes:

[0035] Obstacle distance penalty calculated according to curvature (from graph_B)

[0036] Distance penalty to the end point (from graph_A)

[0037] Reward for vehicle heading close to the narrow point when the current point is near the narrow point.

[0038] In one embodiment, upon detection of a narrow area, the angle sampling density during path expansion is automatically increased, and the sampling strategy is optimized to enhance the reachability and continuity of the path.

[0039] Compared with the prior art, the path search method based on dynamic programming mapping expansion provided by the present application has the following remarkable and unexpected technical effects:

[0040] 1. The search success rate is significantly improved, breaking through the path planning blind area:

[0041] By introducing a dynamic resolution control mechanism, the system can automatically increase the angle sampling density of path expansion in narrow areas, effectively avoiding the phenomenon of path interruption or jamming in narrow spaces in traditional search algorithms. Experiments show that in a typical 5-meter inner wall gap environment, the planning success rate is increased from 68% of the original scheme to more than 97%.

[0042] 2. The planning speed is improved, realizing real-time and efficient search:

[0043] The combination of the end point L1 cost map and the obstacle Vino map double-guiding mechanism makes the search more directional and obstacle avoidance capable. Compared with the traditional pure heuristic A* or RRT expansion method, the average search time is shortened by 42%, and the single path planning time is stably controlled within 150ms, having the ability of real-time operation of vehicle-mounted systems.

[0044] 3. Path quality is significantly optimized, supporting low-speed precise control:

[0045] The composite cost function provided by the application not only considers path length and obstacle distance, but also introduces gear switching penalty, front wheel rotation angle continuity control and other kinematics cost items, so that the generated path has higher executability and smoothness, and is particularly suitable for the discontinuous path tracking model of a low-speed sweeper.

[0046] 4. Enhanced end-point attitude control capability, reduced reversing and escape:

[0047] By applying a heading alignment reward term to the narrow passage target point, the vehicle can automatically adjust to the optimal attitude when approaching the end point or the narrow passage entrance, greatly reducing the probability of abnormal control behaviors such as reversing and secondary adjustment at the end of the path, significantly improving the passage efficiency and enhancing the overall stability of the system.

[0048] 5. Completely autonomous narrow passage structure recognition mechanism, independent of perception:

[0049] The narrow passage point detection algorithm of the application is based on grid sparsity statistical calculation and does not depend on upstream perception modules or semantic maps, has stronger robustness and system decoupling, can be easily transplanted to various unmanned platforms, and is particularly suitable for scenes with incomplete structure information or high real-time requirements.

[0050] 6. General applicability and embedded adaptability, more flexible engineering deployment:

[0051] The algorithm of the application has been deployed on an ARM architecture SoC chip, with running memory control within 40MB, and has stable running ability of low-power and low-frequency devices, greatly reducing the overall deployment threshold of the autonomous driving system, and supporting large-scale industrialization landing.

[0052] 7. Novel technical combination structure, rich parameter tuning space:

[0053] The parameter items (such as alpha, beta, gamma, c_gear, c_phi) in the cost function support dynamic configuration according to task targets, taking into account passage safety, path efficiency and control redundancy requirements, realizing high configurability and strategy flexibility of the path planning system.

[0054] In summary, the application breaks through the adaptability bottleneck of existing path planning algorithms in complex narrow passage scenes through multi-level, multi-target and multi-scale path search modeling strategies, has verifiable system performance improvement and algorithm innovation, and can provide stable and efficient path generation support for future low-speed unmanned sweepers, park mobile robots, garage automatic parking systems and the like. BRIEF DESCRIPTION OF DRAWINGS

[0055] The application will be further described below in conjunction with the accompanying drawings and embodiments.

[0056] Fig. 1 A flow chart of a path planning method of an automatic driving cleaning vehicle in a complex narrow lane scene provided by the application;

[0057] Fig. 2 A flow principle diagram of a path planning method of an automatic driving cleaning vehicle in a complex narrow lane scene provided by the application;

[0058] Fig. 3 An actual effect display diagram of a path planning method of an automatic driving cleaning vehicle in a complex narrow lane scene provided by the application. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions and effects of the application clearer and more explicit, the application will be further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the application and are not used to limit the application. The embodiments of the application will be introduced below in conjunction with the accompanying drawings.

[0060] A path planning method of an automatic driving cleaning vehicle in a complex narrow lane scene provided by the embodiment of the application, please refer to Figs. 1-3 , the method comprises:

[0061] S1, a two-dimensional grid map containing obstacles is constructed, and a grid resolution is set;

[0062] S2, a global cost map guided by an end point is generated based on breadth-first search;

[0063] S3, a maze graph is constructed by obstacle distance transformation;

[0064] S4, a preliminary path is generated by using A* algorithm;

[0065] S5, a narrow lane area meeting a preset condition is extracted based on the path;

[0066] S6, path expansion is performed in the narrow lane area by using an expansion model, and the expansion model is a bicycle model considering a front wheel turning angle;

[0067] S7, a total cost of the path is calculated according to a composite cost function, wherein the total cost comprises a path distance cost, a gear shifting penalty, a turning shifting penalty and an obstacle distance cost;

[0068] S8, the path expansion resolution is dynamically improved in the narrow lane area, and an optimal path is output.

[0069] Specifically, the embodiment of the present application provides a method suitable for path planning of an automatic driving cleaning vehicle in a complex narrow lane scene, and the overall process of the method comprises grid map construction, cost map generation, Voronoi diagram calculation, path preliminary search, narrow lane identification, path expansion, cost evaluation, dynamic sampling adjustment and final path selection, aiming to improve the overall path generation speed while ensuring traffic safety and trajectory feasibility.

[0070] To realize efficient, robust and practical path generation capability, the system first initializes a two-dimensional grid map, and discretizes the operation area into M*N grid units according to the set spatial resolution Δx and Δy, and each unit represents an area in the actual physical space. When the map is initialized, the known static obstacles are mapped as non-passable areas, marked with 1; the rest is air space, marked with 0. This processing ensures the basic expression ability of obstacle avoidance and path planning in the subsequent algorithm.

[0071] After completing the initialization of the grid map, the system uses the breadth-first search (BFS) algorithm to calculate the cost value of each passable grid from the end point. The cost function uses the L1 norm, that is,

[0072] C_{\text{L1}}(i,j)=|i-i_{\text{end}}|+|j-j_{\text{end}}|

[0073] This provides a basic heuristic cost for global path planning, which is simple to calculate, converges quickly, and can effectively provide global heuristic information for path search. This cost map not only provides directional guidance for path expansion, but also is embedded in the subsequent path scoring system as part of the heuristic function.

[0074] After the cost map is calculated, the system further calculates the Euclidean distance DobsDobs of each grid point to its nearest obstacle,

[0075] D_{\text{obs}}(i,j)=\min_{(i',j')\in\text{Obstacles}}\sqrt{(i-i')^2+(j-j')^2}

[0076] The distance transformation algorithm is used to calculate the distance of each grid to the nearest obstacle, which is used for subsequent cost calculation.

[0077] This graph is constructed by the distance transformation algorithm, and provides a spatial reference for the obstacle avoidance cost item in path search. The introduction of Euclidean distance effectively enhances the obstacle avoidance ability of path search in complex space, especially in obstacle-dense or complex-shaped areas, improving the robustness of the algorithm.

[0078] Next, the system performs a preliminary path search between the start and end points by using the heuristic A* algorithm. In this process, the heuristic function f(n) combines C_L1 and D_obs to achieve initial path planning that takes into account both path length and safety. The scoring function is in the form of:

[0079] f(n) = C_L1(n) + aD_obs(n);

[0080] where C_L1 is the end point cost, D_obs is the obstacle distance, a is a weight coefficient, and the balance between safety and efficiency is dynamically adjusted according to the task requirements.

[0081] After the preliminary path Pa is generated, the system performs path structure analysis to extract a set of narrow points Va from the path sequence. The extraction criteria for narrow points are as follows: taking the path point as the center, the number of passable grids in the neighborhood range a x a is investigated, and if the number is less than 1 / 3 of the total number, the point is judged to be the center of the narrow. This method does not rely on external perception results and has high generalization, which can be applied to various unstructured narrow scenarios such as guardrail gaps, roadblock stacks, and offset parking space gaps.

[0082] After identifying the narrow area, the system begins to perform a high-precision path expansion mechanism. Each grid node n is represented as n = {x, y, θ, v, φ}, where x and y are position coordinates, θ is the heading angle, v is the current speed, and φ is the front wheel steering angle. Path expansion is based on a discrete-time model, which uses a bicycle dynamics model to predict the next state n_next: x_next = x + vΔt cos(θ + φ) y_next = y + vΔt sin(θ + φ) θ_next = θ + (vΔt / L) tan(φ) where L is the vehicle wheelbase and Δt is the expansion time step. The above expansion method can generate multiple sub-nodes simultaneously, constructing a path candidate tree and entering the next round of screening.

[0083] For each n_next node generated, the system performs collision detection using the K grid points covered by the vehicle shape model to determine whether it hits the obstacle region (i.e., the grid value is 1). If it hits, the node is discarded directly to ensure path safety.

[0084] Qualified nodes will be scored according to a composite cost function: f(n) = g(n) + ηh(n), where η is the heuristic weight factor, g(n) is the cumulative cost, and h(n) is the heuristic cost. g(n) is accumulated from the following sub-items:

[0085] Δg_dist: the Euclidean distance from the parent node to the current node;

[0086] Δg_gear: gear shift penalty (from forward to reverse or vice versa);

[0087] Δg_phi: steering penalty when front wheel angle is switched from 0 to non-0, or left to right.

[0088] where: Δg_gear = c_gear when gear shift occurs; otherwise, 0. Δg_phi contains two components: c_φ0 when φ_prev is 0 and φ≠0; c_φLR when φ_prev·φ<0, i.e., left to right switch.

[0089] The cost of h(n) is composed as follows:

[0090] Obstacle distance penalty from graph_B, calculated according to curvature

[0091] Distance to goal penalty from graph_A

[0092] Reward when current node is near a narrow point, and vehicle heading is close to the narrow point.

[0093] h_obs = β / (D_obs(n)+ε), obstacle avoidance cost, β controls the strength of obstacle avoidance,

[0094] ε = 0.1 to prevent division by zero

[0095] h_goal = γ·C_L1(n), distance to goal cost, γ is the weight of goal guidance

[0096] h_narrow = k_reward·cos(θ-θ_narrow), when n belongs to the set of narrow points Va, reward for aligning the heading with the direction of the narrow point.

[0097] This cost function design integrates vehicle motion constraints, path smoothness, obstacle safety distance, and pose adaptability, and has high industrial practicality.

[0098] To further enhance the system's ability to pass through narrow areas, the system introduces a dynamic resolution sampling mechanism. When the current expansion point belongs to the narrow area Va, the angular sampling precision Δφ is automatically reduced from the default value to a smaller value φ_min when expanding, to increase the number of path branches and improve the pass rate and trajectory quality.

[0099] Finally, the open_set is managed by a priority queue, and the node with the smallest f(n) is taken out as the current_node in each round. The next_node is expanded and screened, and the close_set is updated. If the next_node is close to the goal area and meets the pose consistency condition (i.e., the angle between θ and θ_goal is less than a threshold), the path is recorded as a feasible solution and output.

[0100] In addition, a penalty balance mechanism is introduced in the path scoring system of the application to impose a penalty on the behavior of driving away from the destination, encouraging the vehicle to maintain a consistent direction when approaching the destination and avoiding path deviation due to reverse driving at the entrance of a narrow lane. The strategy calculates the deviation angle through directional projection and evaluates it comprehensively by combining the L1 cost gradient weight.

[0101] In the implementation process, all weight items (alpha, beta, gamma, eta, c_gear, c_phi0, c_phiLR, k_reward, etc.) are optimized through simulation platform and real vehicle testing to ensure stable performance under different working conditions.

[0102] The real vehicle verification results show that in various typical narrow lane environments such as simulated blocks, underground passages, and obstacle stacking paths, the method proposed by the application can achieve an average path generation time of <150ms and a successful passing rate of more than 98%. Compared with traditional path optimization methods, the planning success rate is improved by more than 30% under challenging conditions such as irregular structure, entrance deviation, and path redundancy.

[0103] In addition, to further improve the universality and engineering deployment efficiency of the algorithm, the system supports a modular deployment structure. The map construction module, cost map generation module, narrow lane detection module, path expansion module, and scoring decision module can be independently deployed and communicate with each other through ROS or other middleware, realizing flexible adaptation to different autonomous driving system platforms.

[0104] The application also introduces a fine-tuning module for dynamic adjustment of the end point, dynamically evaluates the optimal deceleration path based on the heading error of the vehicle's current position and the end point, the angle between the speed vector direction and the channel main axis, and activates the path re-planning module when necessary to avoid end collision.

[0105] In terms of embedded platform transplantation, the method has been implemented on an ARM architecture SoC, supporting platforms with a running frequency as low as 800MHz, with a unit frame search time controlled within 200ms and a memory occupancy of less than 40MB, making it feasible for large-scale deployment.

[0106] In further embodiments, the construction of the grid map discretizes the continuous space into a two-dimensional matrix, with obstacle regions marked as 1 and passable regions marked as 0.

[0107] Specifically, in the grid map construction phase, the system divides the two-dimensional continuous space into a discrete matrix with resolution Δx and Δy, with each grid corresponding to a certain area in the actual physical space. Obstacles are marked as obstacle regions (value 1) through perception modules or map preloading, and the remaining areas are marked as passable regions (value 0). This map serves as the basis for path search and supports path planning and collision detection.

[0108] In a further embodiment, the global cost map calculates the cost value of each grid to the target end point based on L1 norm, for guiding the path search direction.

[0109] In the cost map generation stage, a breadth-first search (BFS) algorithm is adopted to calculate the cost value of each grid based on L1 norm (Manhattan distance) from the target end point. This cost map can effectively guide the path search direction and provide end point heuristic, reducing the risk of search blind area.

[0110] In a further embodiment, the wino map is based on the Euclidean distance of each grid to the obstacle, which is used for obstacle avoidance cost design.

[0111] The generation of the wino map is achieved by calculating the Euclidean distance D_obs(i,j) of each grid point to the nearest obstacle, to construct a local obstacle avoidance structure map. The Euclidean distance information is subsequently used for obstacle avoidance cost item calculation and path smoothness adjustment.

[0112] In a further embodiment, the extraction of the narrow lane region is based on the neighborhood sparsity of the grids in the path, and if the number of passable grids in the neighborhood is lower than a preset threshold, it is determined as a narrow lane.

[0113] After the path search is completed, each node in the preliminary path is determined as a narrow lane point. Taking the current node as the center, the neighborhood grids in the a×a range are taken, and if the number of passable grids is less than 1 / 3 of the total number, the node is marked as a narrow lane point. This determination logic realizes the automatic identification of the narrow lane region, avoiding manual intervention or high dependence on the perception module.

[0114] In a further embodiment, the expansion model includes the lateral position, longitudinal position, heading angle, speed and front wheel steering angle of the vehicle, and motion prediction is performed based on discrete time steps.

[0115] The path expansion stage adopts a five-tuple state node defined by the bicycle model: n={x,y,θ,v,φ}, where x and y are the vehicle positions, θ is the heading angle, v is the speed, and φ is the front wheel steering angle. Under a given time step Δt and wheelbase L, the system predicts the next node state based on the following formulas: x_next=x+v·Δt·cos(θ+φ) y_next=y+v·Δt·sin(θ+φ) θ_next=θ+(v·Δt / L)·tan(φ) This model can truly restore the feasible motion trajectory of the vehicle in low-speed narrow scenarios.

[0116] In a further embodiment, when a narrow lane region is detected, the angle sampling density during path expansion is automatically increased, and the sampling strategy is optimized to enhance the reachability and continuity of the path.

[0117] When an expansion node is detected in a narrow lane area (i.e., a point in the Va set), the system automatically reduces the angular sampling interval Δφ, improves path diversity, and ensures that the algorithm can search for a feasible path in a scene that is extremely narrow but still passable.

[0118] During the expansion process, the open_set priority queue is sorted by the total cost f(n), and the system selects the node with the minimum total cost current_node in each round to expand, and adds the expansion result to the open_set, and the current node is added to the close_set. This cycle is iterated until a node meets the end point proximity judgment condition and the heading angle consistency, and then the search is considered successful.

[0119] The method supports modular deployment, combines a map service module, a path candidate generation module, a narrow lane detection module, a dynamic expansion module and a path scoring module, and is integrally deployed with a middleware such as ROS, is suitable for various heterogeneous automatic driving platforms, and realizes an engineering path planning capability that can be landed, reused and optimized.

[0120] In summary, the present application proposes a path search method combining global heuristic and local structure perception, which constructs a path planning framework with search efficiency, reliable passage and attitude adaptability through dynamic programming mapping expansion and multi-source cost function fusion. The method fully considers the kinematic constraints, environmental uncertainty and control response requirements of the automatic driving sweeper vehicle in a narrow and complex scene, and breaks through the typical technical bottlenecks of the traditional planning scheme, such as search failure, unstable attitude and uncontrollable path in a narrow lane environment.

[0121] Each technical measure in the present application, whether it is to extract narrow lane points based on grid sparseness, to introduce a heading alignment reward mechanism in path scoring, or to integrate a dynamic sampling density adjustment strategy in the expansion model, has undergone systematic structural design and parameter optimization, and can realize end-to-end adaptive path search and optimization without relying on high-precision maps and semantic inputs. The scheme has high universality and good deployability, and is suitable for various low-speed autonomous mobile systems, especially in the fields of sweepers, park delivery vehicles, underground parking robots and the like.

[0122] The present application not only has good engineering landing properties, but also has theoretical innovation points, including a dynamic resolution control mechanism, a multi-map fusion guidance strategy, a path adaptability mechanism for structural uncertainty, and a heading reward driven end point attitude optimization strategy, which can be widely applied to low-speed autonomous driving systems, multi-sensor fusion navigation platforms and future urban mobile robot autonomous passage and other key fields.

[0123] The path search method has high module decoupling and system coupling capabilities, combines structure data driving and kinematic constraint control strategies, breaks through the efficiency and effect bottleneck of traditional methods in narrow channel environment, and shows a cross-scene universal technology route, which is a key path planning component worthy of promotion in the construction of future unmanned driving systems.

[0124] The various embodiments and technical details described in the specification serve to fully support and disclose the claimed path search method, which is not limited to specific parameter configurations and system implementation methods. Any equivalent changes and improvements made on the basis of the technical idea of the present application are within the scope of the present application.

Claims

1. A method for path planning of an autonomous driving cleaning vehicle in a complex narrow lane scene, characterized in that, The method comprises: S1, constructing a two-dimensional grid map containing obstacles, and setting a grid resolution; S2, generating a global cost map guided by the end point based on breadth-first search; S3, constructing a Viterbi graph through obstacle distance transformation; S4, generating a preliminary path using A* algorithm; S5, extracting a narrow lane area meeting a preset condition based on the path; S6, expanding the path in the narrow lane area using an expansion model, which is a bicycle model considering the front wheel steering angle; S7, calculating the total cost of the path according to a composite cost function, wherein the total cost includes path distance cost, gear shifting penalty, steering shifting penalty and obstacle distance cost; S8, dynamically improving the path expansion resolution in the narrow lane area, and outputting an optimal path.

2. The path planning method for an autonomous driving cleaning vehicle in a complex narrow lane scene according to claim 1, wherein, The construction of the grid map discretizes the continuous space into a two-dimensional matrix, and the obstacle area is marked as 1 and the passable area is marked as 0.

3. The path planning method for an autonomous driving cleaning vehicle in a complex narrow lane scenario according to claim 1, wherein, The global cost map calculates the cost value of each grid to the target end point based on L1 norm, which is used to guide the path search direction.

4. The path planning method for an autonomous driving cleaning vehicle in a complex narrow lane scenario according to claim 1, wherein, The Viterbi graph is calculated based on the Euclidean distance of each grid to the obstacle, and the distance is used for the cost design of obstacle avoidance.

5. The path planning method for an autonomous driving cleaning vehicle in a complex narrow lane scenario according to claim 1, wherein, The scoring function of the A* algorithm is: f(n)=C_{\text{L1}}(n)+\alpha\cdot D_{\text{obs}}(n); Wherein C_{\text{L1}} is the end point cost, D_{\text{obs}} is the obstacle distance, and α is the weight coefficient.

6. The path planning method for an autonomous driving cleaning vehicle in a complex narrow lane scenario according to claim 1, wherein, The extraction of the narrow lane area is based on the neighborhood sparsity of the grid in the path. If the number of passable grids in the neighborhood is lower than the preset threshold, it is determined as a narrow lane.

7. The path planning method for an autonomous driving cleaning vehicle in a complex narrow lane scene according to claim 1, wherein, The expansion model includes the lateral position, longitudinal position, heading angle, speed and front wheel steering angle of the vehicle, and motion prediction is performed based on discrete time steps.

8. The path planning method for an autonomous driving cleaning vehicle in a complex narrow lane scene according to claim 1, wherein, The total cost function f(n) includes: Euclidean distance cost between parent nodes; Penalty term for gear shifting; Penalty term for front wheel steering angle switching from 0 to non-zero; Penalty term for left and right switching of the front wheel. 9.The path planning method of the autonomous driving cleaning vehicle in a complex narrow lane scene according to claim 1, wherein, The heuristic cost h(n) includes: Obstacle distance penalty according to curvature (obtained from graph_B) Distance penalty to the end point (obtained from graph_A) When the current point is near the narrow lane point, the vehicle heading is close to the narrow lane point.

10. The method of claim 1, wherein the autonomous driving cleaning vehicle in a complex narrow lane scene is planned a path. When the narrow lane area is detected, the angle sampling density during path expansion is automatically improved, and the sampling strategy is optimized to enhance the accessibility and continuity of the path.

Images