A global planning method for structured road cleaning operation of unmanned sweeper

Abstract

The application discloses a kind of global planning methods for the structured road cleaning operation of unmanned cleaning vehicle, and the application relates to unmanned cleaning vehicle path planning technical field, and the method comprises the following steps: S1.building topological graph and establishing cost model;S2.based on the cleaning strategy initialization of task mode;S3.based on the topological graph and cost model constructed in step S1, construct the adjacent matrix containing auxiliary node, the total cost containing auxiliary node is approximately optimal closed loop circuit by algorithm solution, remove auxiliary node, and generate an optimal lane access sequence starting and ending with starting point according to the actual starting position of vehicle;S4.iterative path generation and splicing, finally output the complete global cleaning path spliced together.This application is suitable for unmanned cleaning vehicle in structured road scene to realize full coverage cleaning, curb cleaning, center cleaning operation, to adapt to different operation requirements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned cleaning vehicle path planning, in particular to a global planning method for structured road cleaning operation of unmanned cleaning vehicle. BACKGROUND

[0002] Currently, the operation path planning of unmanned cleaning vehicle on structured road usually relies on preset road sequence template or mode of driving along fixed lane line. For example, the vehicle needs to be pre-configured with the road sequence that needs to be cleaned, or simply drives back and forth along a certain lane center line. This method has the following main shortcomings:

[0003] 1. Poor flexibility: it is difficult to meet the diversified needs of different users for cleaning modes (such as edge cleaning, full coverage cleaning, and center cleaning), and each time the task is changed or the map is updated, the path template may need to be reconfigured manually, which is not user-friendly for ordinary users.

[0004] 2. Low intelligence: the existing scheme usually does not have global optimization capability, which may generate a large number of repeated paths, empty paths or turns, resulting in low operation efficiency, increased energy consumption and other problems.

[0005] 3. Non-optimal path: the existing global path planning method is prone to generate non-optimal routes, and the abstraction of physical world road is not comprehensive enough, making it difficult to achieve truly optimal path planning in real scenarios. SUMMARY

[0006] In view of the poor flexibility, low intelligence and non-optimal path of the path planning of unmanned cleaning vehicle in the prior art, the present application provides a global planning method for structured road cleaning operation of unmanned cleaning vehicle, which is suitable for full coverage cleaning, edge cleaning and center cleaning operation of unmanned cleaning vehicle in structured road scene to adapt to different operation requirements.

[0007] The purpose of the present application is achieved by the following technical solution: a global planning method for structured road cleaning operation of unmanned cleaning vehicle, comprising the following steps:

[0008] S1. Constructing a topological graph and establishing a cost model; load the map file to obtain road and lane information, and construct a road network topological graph composed of nodes and directed edges by taking the lane as the basic planning unit; and calculate the node cost and edge cost by a cost function to obtain the path cost;

[0009] S2. Initialize the cleaning strategy based on the task mode;

[0010] S3. Based on the topology graph and cost model constructed in step S1, the shortest path cost between any two nodes to be cleaned is calculated, an adjacency matrix containing auxiliary nodes is constructed, and a closed loop circuit containing auxiliary nodes with approximately optimal total cost is solved by an algorithm, and an optimal lane access sequence starting and ending at the actual starting point of the vehicle is generated after removing the auxiliary nodes;

[0011] S4. Iterative path generation and splicing, finally output the complete global cleaning path spliced.

[0012] Specifically, the node cost is mainly determined by the lane length and cleaning state, and a high penalty value is applied to the cleaned lane to reduce its priority in subsequent planning; the edge cost includes straight cost and left / right lane change cost, the penalty value of straight cost is the lowest, and the penalty value of left / right lane change cost is higher than that of straight cost, to encourage the algorithm to generate a good continuity path and reduce unnecessary lane change and turning.

[0013] Further, the formula of the node cost is: cost_node = length + P_visited is_visited, wherein length is the total length of the lane, P_visited is the penalty value of the lane that has completed cleaning, and is_visited is a mark value of whether the cleaning is completed; the formula of the edge cost is: cost_edge = P_straight is_straight + P_left is_left + P_right is_right, P_straight is the straight penalty value, P_left is the left turn cost, P_right is the right turn cost, is_straight, is_left and is_right are mark values of whether the connection is straight, left turn and right turn; the path cost is the sum of all node costs and edge costs on the path from the starting point to the target point.

[0014] Specifically, the step S4 adopts an iterative planning process, which includes the following steps:

[0015] S41. Single sequence path generation: for the optimal lane access sequence obtained in step S3, each pair of adjacent nodes is processed in turn; if the two nodes are directly connected in the topology graph, the lane reference line such as the edge line or the center line is directly used; if not directly connected, a high-precision map path planner is called to generate the shortest feasible navigation path between the two points, so as to splice the paths of cleaning segments and transition segments in order;

[0016] S42. State update and iterative judgment: after completing a global sequence planning, update the cleaning state of all visited lanes; check whether all to-be-cleaned sides of all lane nodes have been planned;

[0017] S43. Iterative loop: if not completed, take the current path endpoint as the new starting point, and construct a new distance matrix with the remaining to-be-cleaned node set, repeat steps S3 and steps S41 and S42 of the planning process, generate the optimal sequence and path of the next cleaning, and splice it to the existing path;

[0018] S44. Termination and output: when all to-be-cleaned tasks are completed, output the final spliced complete global cleaning path.

[0019] Further, the step S3 uses the LKH algorithm to solve to obtain the optimal cleaning road sequence starting from the vehicle starting point; the connection cost of the auxiliary node and all real nodes is set to 0, and the self-connection cost of the auxiliary node is set to a maximum value; After solving the algorithm containing the auxiliary node closed loop, the auxiliary node index is removed from the loop sequence to obtain a sequence containing only real nodes.

[0020] Specifically, the step S3 is, first, based on the topology graph and cost model constructed in step S1, using A algorithm, forming a q×q asymmetric adjacency matrix, where q is the total number of to-be-cleaned nodes; then, add a row and a column as an auxiliary node in the adjacency matrix to construct an expanded adjacency matrix with a size of (q+1)×(q+1); input the adjacency matrix into the LKH solver, and the LKH algorithm obtains a Hamiltonian circuit that visits all nodes and returns to the starting point through an efficient K-opt edge exchange and local search strategy, and the total cost of the circuit is approximately optimal; finally, after removing the auxiliary node from the loop, the closed loop output by the LKH algorithm is generated into an open loop optimal lane visiting sequence starting and ending at the starting point according to the actual starting position of the vehicle.

[0021] Specifically, the step S3, based on the optimal cleaning road sequence pseudo code of the LKH algorithm, has the following specific content:

[0022] Input: high-precision map hdmap, structured road topology graph, vehicle starting point s, one-key cleaning task mode;

[0023] Output: the optimal cleaning road sequence starting from the vehicle starting point s;

[0024] Including the following steps:

[0025] S31. Set the to-be-cleaned attribute of the lane at initialization according to the one-key cleaning task mode, and determine the cleaning number according to the road width;

[0026] S32. Calculate the cost value between any two road nodes m and n according to the structured road topology and path cost value, and construct the corresponding adjacency matrix Adj(q x q), where the size q is the number of structured road nodes to be cleaned;

[0027] S33. Add an auxiliary node to the adjacency matrix Adj, and construct an extended adjacency matrix Adj_extended((q+1) x (q+1)), where the connection cost of the auxiliary node with all real nodes is 0;

[0028] S34. Call the LKH algorithm to obtain the optimal loop A_extended[n1, n2,..., np, dummy_node, n1] containing the auxiliary node based on the extended adjacency matrix;

[0029] S35. Remove the auxiliary node dummy_node from A_extended to obtain a closed loop A[n1, n2,..., np, n1] containing only real nodes;

[0030] S36. According to the starting point s of the vehicle, match the corresponding lane node lane_id, and find the corresponding position index according to the loop output in step S35, and finally generate the optimal cleaning road sequence A[index:end, 1:index-1];

[0031] S37. Based on the road boundary and center information provided by the high-precision map hdmap, the vehicle assembles the cleaning path along the optimal cleaning road sequence from the current position;

[0032] S38. Query whether all roads meet the cleaning of the road edge to be cleaned, and if all roads have been cleaned, output the final cleaning route; if not all roads have been cleaned, update the adjacency matrix, where the cleaned roads do not participate in the next round of loop, and continue to execute steps S32 to S37, and after the execution is completed, splice the path result of the last round.

[0033] Further, it further includes step S5 data persistence and reuse, which includes planning result caching and fast reuse; the planning result caching is that, in the first or previous planning process, the system serializes and stores the optimal path and its cost between any two lane nodes under different cleaning modes; the fast reuse is that, in subsequent planning tasks, if it is judged that the starting point, ending point and map have not changed, the pre-calculated path and cost value can be directly read from the cache without re-calling the expensive path planning algorithm.

[0034] Further, the step S3 uses a genetic algorithm, a simulated annealing algorithm or an ant colony algorithm for solution.

[0035] Specifically, step S2 involves initializing the cleaning attributes of all lane nodes according to the cleaning task mode selected by the user before planning begins; the task mode includes edge cleaning mode, full coverage cleaning mode, and center cleaning mode.

[0036] Compared with existing technologies, this invention can automatically and efficiently generate a global reference path that starts from any starting point, covers all areas to be cleaned, and has an approximately optimal total cost (such as driving distance, number of lane changes, number of turns, etc.) based on the cleaning mode selected by the user (edge, full coverage, center), thereby improving the automation level and work efficiency of cleaning operations. Attached Figure Description

[0037] Figure 1 This is a structured panoramic view of a road in a real-world scenario, as described in an embodiment of the present invention.

[0038] Figure 2 This is a topology map corresponding to the structured road in a real-world scenario according to an embodiment of the present invention.

[0039] Figure 3 for Figure 1 The actual scenario of full-coverage cleaning planning routes for structured roads.

[0040] Figure 4 for Figure 1 The actual scenario of structured road edge cleaning route.

[0041] Figure 5 for Figure 1 The actual scenario of a structured road with a centered cleaning route. Detailed Implementation

[0042] To facilitate understanding of the present invention, the technical solutions and advantages of the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Furthermore, any of the technical features mentioned below (including implicit or disclosed features), as well as any technical feature directly shown or implied in the figures, can be arbitrarily combined or deleted among these technical features to form other embodiments that may not be directly or indirectly mentioned in the present invention. The accompanying drawings show preferred embodiments of the present invention. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention.

[0043] This invention provides a hierarchical, iterative global cleaning path planning method. Its core lies in decomposing the complex structured road cleaning task into three levels of problems, and solving and iterating layer by layer. These are:

[0044] Lane-level cleaning mode decision layer: according to the cleaning task type (such as edge, full coverage or center) and lane geometry information, the cleaning mode of each lane (such as cleaning which side or several sides) and the required number of times are determined.

[0045] Lane sequence global optimization layer: abstract the "access all to-be-cleaned lane nodes" as a traveling salesman problem (TSP), and solve the globally approximate optimal lane access sequence by using an improved LKH heuristic algorithm.

[0046] Lane transition and path generation layer: for adjacent but not directly connected lane nodes in the optimized sequence, a mature path planning algorithm is called to generate a specific and drivable transition navigation path, which is seamlessly spliced with the cleaning trajectory (such as the edge line or center line) in the lane.

[0047] The whole method uses an iterative planning mechanism in full coverage mode: when a single global sequence planning cannot complete the multi-pass cleaning of a wide lane at one time, the state of the cleaned lane is updated, and the above three layers of planning and splicing are performed again based on the remaining to-be-cleaned area until all lane areas are covered.

[0048] Specifically, the global planning method for structured road cleaning operation of the unmanned cleaning vehicle of the present application comprises the following steps:

[0049] S1. Problem modeling and data structure construction.

[0050] Specifically, step S1 specifically includes two parts of topological graph construction and cost model establishment.

[0051] Topological graph construction refers to loading a high-precision map file hdmap.xml to obtain road (Road) and lane (Lane) information. A directed topological graph is constructed with lane (Lane) as the basic planning unit, and the obtained topological graph constitutes a complete road network topological graph composed of nodes (Node) and directed edges (Edge), as shown in Figure 1 and Figure 2 . Figure 1 For an example of an actual scenario structured road panoramic map, Figure 2 is the topological graph corresponding to the actual scenario structured road, Figure 2 in which a plurality of lane nodes lane_id are marked, wherein lane represents the road number, and id represents the identification of the lane. Taking lane node 9_2 as an example, it represents the 2nd lane of the road numbered 9. Figure 2 The lane node, direct edge and left-right lane-changing edge are clearly shown, and the connectivity between lanes is clearly expressed.

[0052] Wherein, the node (Node) model: each lane (Lane) is defined as a node (Node), and attributes are as shown in Table 1, for recording the geometric information, connection relationship, cleaning state and cost of the lane.

[0053] The edge (Edge) model: the directed connection relationship between lanes is defined as an edge (Edge), and attributes are as shown in Table 2, for recording the direction of connection and switching cost.

[0054]

[0055] Table 1 Node lane information

[0056]

[0057] Table 2 Edge lane connection edge information

[0058] The cost model is established, that is, for the guide algorithm to generate efficient and continuous cleaning path, a comprehensive cost function is designed. The core idea of the application is to establish a corresponding model for complex multi-road cleaning planning, design a reasonable path cost, use a search type algorithm or an intelligent type optimization algorithm to calculate the optimal sequence of the optimal lane node, and combine the basic lane boundary and center line information of the high-precision map to generate the corresponding edge and full coverage path. The cost of the path can be composed of the cost value of each node (Node) and connection edge (Edge) in the path, that is, the node cost and the edge cost.

[0059] Node cost: mainly determined by lane length and cleaning state. A high penalty value is applied to the lane that has been cleaned, to reduce its priority in subsequent planning, and the formula can be simplified as: cost_node = length + P_visited is_visited, wherein, length is the total length of the lane, P_visited is the penalty value for the lane that has completed cleaning, and the value is a large enough penalty value, and in the application, P_visited is set to 1000.0, and is_visited is a mark value (0 / 1) whether the cleaning is completed.

[0060] Edge cost: for punishing non-ideal planning behavior. The straight cost is the lowest, and the left / right lane changing cost is higher than the straight cost, to encourage the algorithm to generate a good continuity path and reduce unnecessary lane changing and turning. The formula can be simplified as: cost_edge = P_straight is_straight + P_left is_left + P_right is_right, wherein P_straight is a straight penalty value, the cost of the unmanned sweeper performing a cleaning task in a straight line on an open road is 0, P_left is the penalty value of left turn, which is set to 50.0, P_right is the penalty value of right turn, which is set to 20.0, the values of P_left and P_rigth are determined according to the corresponding national legal driving side, for the case of legal right-hand driving, the cost of left turn is higher than that of right turn; for the case of legal left-hand driving, the cost of right turn is higher than that of left turn, is_straight, is_left and is_right are flag values (0 / 1) of straight connection, left turn and right turn.

[0061] Path cost: the sum of the cost of all nodes and the cost of edges on the path from the starting point to the target point.

[0062] S2. Initialization of cleaning strategy based on task mode.

[0063] Specifically, before the planning starts, the method initializes the to-be-cleaned attribute of all lane nodes according to the cleaning task mode selected by the user:

[0064] Edge cleaning mode: the method identifies the boundary lane in the road network, such as the leftmost or rightmost lane, and assigns a corresponding single-edge cleaning task, such as left-edge or right-edge, to it.

[0065] Full coverage cleaning mode: the method dynamically calculates the number of passes required for each lane according to the lane width and the effective width of the sweeper, for example, a wide lane may require three passes of left, middle and right, and creates virtual "cleaning task nodes" for each cleaning task. These nodes share the same physical lane, but have independent cleaning states, such as "left side to be cleaned", "middle to be cleaned".

[0066] Center cleaning mode: all lane nodes are marked as only needing to be cleaned along the center line of the lane.

[0067] S3. Global lane sequence optimization based on LKH algorithm.

[0068] Specifically, the problem of "visiting all to-be-cleaned lane nodes or virtual cleaning task nodes and returning to the starting point" is accurately abstracted as an asymmetric traveling salesman problem, i.e. ATSP. The present invention uses the LKH (full name Lin-Kernighan-Helsgaun) algorithm to solve it to obtain the optimal cleaning road sequence starting from the vehicle starting point s.

[0069] Distance matrix construction: based on the topological graph and cost model constructed in step S1, the shortest path cost between any two to-be-cleaned nodes is calculated. A or Dijkstra algorithm can be used, and the present invention uses A The algorithm forms a q×q asymmetric adjacency matrix, where q is the total number of nodes to be cleaned. A row and column are added to the adjacency matrix as auxiliary nodes, constructing an extended adjacency matrix of size (q+1)×(q+1), thus forming an adjacency matrix containing the auxiliary nodes.

[0070] LKH Solution: Input the above adjacency matrix into the LKH solver. The LKH algorithm quickly finds a Hamiltonian cycle that visits all nodes, including auxiliary nodes, and returns to the starting point through efficient K-opt edge swapping and local search strategies. The total cost of this cycle is approximately optimal.

[0071] Sequence adaptation starting point: After removing the auxiliary node from the loop, the closed loop output by the LKH algorithm is "cut off" and reassembled according to the actual starting position of the vehicle to generate an open-loop optimal lane access sequence with the starting point as the beginning and end.

[0072] In this sequence, the connection cost between the auxiliary node and all real nodes is set to 0, and the self-connection cost of the auxiliary node is set to a maximum value. After solving the closed loop containing the auxiliary node through the algorithm, the auxiliary node index is removed from the loop sequence to obtain a sequence containing only real nodes.

[0073] S4. Iterative path generation and splicing.

[0074] Specifically, since a single lane may require multiple sweeps in the full coverage mode, step S4 adopts an iterative planning process, which includes the following steps:

[0075] S41. Single Sequence Path Generation: For the optimal lane sequence obtained in step S3, process each pair of adjacent nodes sequentially. If the two nodes are directly connected on the topology map, use the lane reference line directly, such as the edge line or center line; if they are not directly connected, call the high-precision map path planner to generate the shortest feasible navigation path between the two points. Then, concatenate these "cleaning segment" and "transition segment" paths in sequence.

[0076] S42. State Update and Iteration Judgment: After completing one global sequence planning, update the cleaning status of all visited lanes. Check whether all cleaning sides of all lane nodes have been planned.

[0077] S43. Iterative Loop: If not completed, take the current path endpoint as the new starting point, construct a new distance matrix with the remaining set of nodes to be cleaned, repeat step S3 and steps S41 and S42 of this planning process, generate the optimal sequence and path for the next cleaning pass, and attach them to the existing path.

[0078] S44. Termination and output: when all the tasks to be cleaned are completed, output the final spliced complete global cleaning path. Each path point in the path has an attribute label, such as "left edge operation" and "turn field driving", for identification and execution by downstream control modules.

[0079] S5. Data persistence and reuse.

[0080] Specifically, step S5 is used for performance optimization. To improve planning efficiency, a data serialization mechanism is designed, including:

[0081] Planning result cache: during the first or previous planning process, the system serializes and stores the optimal path and its cost between any two lane nodes under different cleaning modes.

[0082] Fast reuse: in subsequent planning tasks, if it is determined that the start point, end point and map have not changed, the pre-computed path and cost value can be directly read from the cache without re-invoking the expensive path planning algorithm, greatly improving the response speed, especially for large-scale road networks.

[0083] Through the above global planning method, an optimal global cleaning planning route for an unmanned cleaning vehicle in a structured road environment can be obtained under a full coverage / edge cleaning, or a central cleaning mode. As shown in Figures 3-5 , Figure 3 is a full coverage cleaning planning route, Figure 4 is an edge cleaning route, Figure 5 is a central cleaning route.

[0084] The above step S3, the pseudo code of the optimal cleaning road sequence based on the LKH algorithm, has the following specific content:

[0085] Input: high-precision map hdmap, structured road topology, vehicle start point s, one-key cleaning task mode (edge / full coverage / central).

[0086] Output: the optimal cleaning road sequence starting from the vehicle start point s.

[0087] Including the following steps:

[0088] S31. Set the cleaning attribute of the lane at the initialization according to the one-key cleaning task mode, such as determining the cleaning number according to the road width, and ensuring at least 20 cm of brush overlap width between adjacent cleaning paths in the same lane to ensure that the brush hard edge does not press the lane line;

[0089] S32. Calculate the cost value between any two road nodes m and n according to the structured road topology and path cost value, and construct the corresponding adjacency matrix Adj(qxq), where the size q is the number of structured road nodes to be cleaned.

[0090] S33. Add an auxiliary node to the adjacency matrix Adj, and construct an extended adjacency matrix Adj _extended((q+1) x (q+1)), wherein the connection cost of the auxiliary node to all real nodes is 0;

[0091] S34. Call the LKH algorithm to obtain an optimal loop A _extended[n1, n2,..., np, dummy_node, n1] containing the auxiliary node based on the extended adjacency matrix;

[0092] S35. Remove the auxiliary node dummy_node from A _extended to obtain a closed loop A[n1, n2,..., np, n1] containing only real nodes;

[0093] S36. Match the corresponding lane node lane_id according to the vehicle starting point s, find the corresponding position index according to the loop output in step S35, and finally generate an optimal sweeping road sequence A[index:end, 1:index-1];

[0094] S37. Based on the road boundary and center information provided by the high-precision map hdmap, the vehicle assembles a sweeping path along the optimal sweeping road sequence from the current position;

[0095] S38. Query whether all roads meet the road edge to be swept and have completed sweeping, if all meet the swept road edge, end the output of the final sweeping route; if not all meet, update the adjacency matrix, wherein the swept road does not participate in the next round of circulation, continue to execute steps S32 to S37, and splice the path result of the last round after the execution is completed.

[0096] The above step S41, in the point-to-point planning, the optimal sweeping lane sequence calculated based on the LKH algorithm may not be adjacent, and therefore corresponding navigation planning to the specified road for sweeping needs to be performed. The high-precision map path planner can apply the global path planning method, electronic device and storage medium of Chinese patent application CN202310629106.7.

[0097] Alternative technical solutions:

[0098] Alternative solving algorithm: the global sequence optimization layer of the present application is not limited to the LKH algorithm, and other meta-heuristic algorithms such as genetic algorithm (GA), simulated annealing algorithm (SA) or ant colony algorithm (ACO) can also be used for solving, in order to adapt to different computing resources and real-time requirements.

[0099] Cost model adaptation: the cost coefficients in the formula (e.g. the lane changing penalty value) can be learned offline or adjusted online based on actual operation data (time, energy consumption), so that the generated path is more in line with the actual operation efficiency indicators.

[0100] The above embodiments are only preferred embodiments of the present application, and cannot be used to limit the scope of protection of the present application. For those skilled in the art, it can be understood that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A global planning method for structured road sweeping operations using unmanned sweepers, characterized in that, Includes the following steps: S1. Constructing the topology graph and establishing the cost model; by loading map files to obtain road and lane information, a road network topology graph consisting of nodes and directed edges is constructed using lanes as the basic planning unit; Furthermore, the path cost is obtained by calculating the node cost and edge cost using a cost function; S2. Initialize the cleaning strategy based on the task mode; S3. Based on the topology and cost model constructed in step S1, calculate the shortest path cost between any two nodes to be cleaned, construct an adjacency matrix containing auxiliary nodes, and solve the closed loop with the total cost of the auxiliary nodes approximately optimal through the algorithm. After removing the auxiliary nodes, generate an optimal lane access sequence with the starting point as the beginning and the end point based on the actual starting position of the vehicle. S4. Iterative path generation and splicing, finally outputting a spliced, complete global cleaning path; Step S4 employs an iterative planning process, which includes the following steps: S41. Single Sequence Path Generation: For the optimal lane access sequence obtained in step S3, process each pair of adjacent nodes in sequence; if the two nodes are directly connected on the topology map, the lane reference line is directly used, which is either the edge line or the center line; if they are not directly connected, the high-precision map path planner is called to generate the shortest feasible navigation path between the two points, thereby splicing the paths of the cleaning section and the transfer section in sequence. S42. State Update and Iteration Judgment: After completing a global sequence planning, update the cleaning status of all visited lanes; check whether all cleaning sides of all lane nodes have been planned. S43. Iterative loop: If not completed, take the current path end point as the new starting point, construct a new distance matrix with the remaining set of nodes to be cleaned, repeat step S3 and steps S41 and S42 of this planning process, generate the optimal sequence and path for the next cleaning pass, and attach them to the existing path; S44. Termination and Output: When all cleaning tasks are completed, output the final, complete global cleaning path.

2. The global planning method for structured road sweeping operations using unmanned sweepers as described in claim 1, characterized in that, The node cost is mainly determined by the lane length and cleaning status. Lanes that have been cleaned are subject to a high penalty value to reduce their priority of being repeatedly selected in subsequent planning. The edge cost includes straight-going cost and left / right lane change cost. The penalty value for straight-going cost is the lowest, and the penalty value for left / right lane change cost is higher than that for straight-going cost, so as to encourage the algorithm to prioritize generating paths with good continuity and reduce unnecessary lane changes and turns.

3. The global planning method for structured road sweeping operations using unmanned sweepers as described in claim 2, characterized in that, The formula for the node cost is: cost_node = length + P_visited is_visited, where length is the total length of the lane, P_visited is the penalty value for lanes that have been cleaned, and is_visited is a flag value indicating whether cleaning is complete; the formula for the edge cost is: cost_edge = P_straight is_straight+ P_left is_left + P_right is_right, P_straight are the direct connection penalty values, P_left is the cost of turning left, P_right is the cost of turning right, and is_straight, is_left, and is_right are the flag values ​​indicating whether the connection is a direct connection, a left turn, or a right turn; the path cost is the sum of the costs of all nodes and edges on the path from the starting point to the target point.

4. The global planning method for structured road sweeping operations using unmanned sweepers as described in claim 1, characterized in that, Step S3 uses the LKH algorithm to solve for the optimal sweeping road sequence starting from the vehicle's starting point; the connection cost between the auxiliary node and all real nodes is set to 0, and the self-connection cost of the auxiliary node is set to the maximum value; after solving the closed loop containing the auxiliary node through the algorithm, the auxiliary node index is removed from the loop sequence to obtain a sequence containing only real nodes.

5. The global planning method for structured road sweeping operations using unmanned sweepers as described in claim 4, characterized in that, Step S3 is as follows: First, based on the topology graph and cost model constructed in step S1, use A... The algorithm first forms a q×q asymmetric adjacency matrix, where q is the total number of nodes to be cleared. Then, a row and a column are added to the adjacency matrix as auxiliary nodes to construct an extended adjacency matrix of size (q+1)×(q+1). The adjacency matrix is ​​then input into the LKH solver. The LKH algorithm obtains a Hamiltonian cycle that visits all nodes and returns to the starting point through efficient K-opt edge swapping and local search strategies. The total cost of this cycle is approximately optimal. Finally, after removing the auxiliary nodes from the cycle, the closed-loop cycle output by the LKH algorithm is used to generate an open-loop optimal lane access sequence with the starting point as the beginning and the end point, based on the actual starting position of the vehicle.

6. The global planning method for structured road sweeping operations using unmanned sweepers as described in claim 5, characterized in that, Step S3, based on the pseudocode of the optimal road cleaning sequence using the LKH algorithm, is as follows: Input: High-precision map (hdmap), structured road topology map, vehicle starting point (s), one-click cleaning task mode; Output: The optimal road cleaning sequence starting from vehicle origin s; Includes the following steps: S31. Set the cleaning attributes of the lane during initialization according to the one-click cleaning task mode, and determine the number of cleaning passes according to the road width; S32. Calculate the cost between any two road nodes m and n based on the structured road topology map and path cost, and construct the corresponding adjacency matrix Adj(q×q); S33. Add auxiliary nodes to the adjacency matrix Adj and construct the extended adjacency matrix Adj_extended((q+1)×(q+1)), where the connection cost between the auxiliary nodes and all real nodes is 0; S34. Call the LKH algorithm to obtain the optimal circuit A_extended[n1,n2,...,np,dummy_node,n1] containing auxiliary nodes based on the extended adjacency matrix; S35. Remove the auxiliary node dummy_node from A_extended to obtain a closed loop A[n1,n2,...,np,n1] containing only real nodes; S36. Match the corresponding lane node lane_id according to the vehicle starting point s, find the corresponding position index according to the loop output in step S35, and finally generate the optimal sweeping road sequence A[index:end,1:index-1]; S37. Based on the road boundary and center information provided by the high-precision map hdmap, the vehicle assembles a cleaning path from its current position along the optimal cleaning road sequence; S38. Check if all roads have been cleaned. If all roads have been cleaned, end the process and output the final cleaning route. If not all roads have been cleaned, update the adjacency matrix. The cleaned roads will not participate in the next round of the loop. Continue to execute steps S32 to S37. After execution, concatenate the path results from the previous round.

7. The global planning method for structured road sweeping operations using unmanned sweepers as described in claim 1, characterized in that, It also includes step S5, data persistence and reuse, which includes planning result caching and fast reuse. The planning result caching means that during the first or subsequent planning processes, the system serializes and stores the optimal path and its cost between any two lane nodes under different cleaning modes. The fast reuse means that in subsequent planning tasks, if it is determined that the starting point, ending point and map have not changed, the pre-calculated path and cost can be directly read from the cache without having to call the expensive path planning algorithm again.

8. The global planning method for structured road sweeping operations using unmanned sweepers as described in claim 1, characterized in that, Step S3 is solved using a genetic algorithm, simulated annealing algorithm, or ant colony algorithm.

9. The global planning method for structured road sweeping operations using unmanned sweepers as described in claim 1, characterized in that, Step S2 is as follows: before the planning begins, the method initializes the cleaning attributes of all lane nodes according to the cleaning task mode selected by the user; the task mode includes edge cleaning mode, full coverage cleaning mode and center cleaning mode.

Images