Subway service robot real-time path planning method based on graph neural network

Abstract

The invention discloses a subway service robot real-time path planning method based on a graph neural network, and the method comprises the following steps: scanning the environment in a subway station, completing the map construction, and obtaining basic environment data; performing target detection and personnel statistics to form an environment and passenger flow fusion map; taking the fused map as input, setting a starting point and a target position, and outputting a global path search initial state; calling an improved A * algorithm to carry out node expansion, calculating a cost estimation function, and generating candidate paths; performing redundant node deletion and arc smoothing processing on the candidate path, performing step length dynamic adjustment, and outputting an optimized global path; performing local path planning by adopting an improved dynamic window method, and outputting an update path in real time; a passenger evacuation scheme is generated and is output through a display screen and voice broadcast, evacuation is guided, and passing is guaranteed. According to the invention, real-time path planning of the subway service robot is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot path planning, and particularly relates to a subway service robot real-time path planning method based on a graph neural network. BACKGROUND

[0002] With the continuous acceleration of urbanization, as an important part of urban public transportation, subway has dense passenger flow in the station and frequent dynamic changes. How to achieve efficient and safe navigation of service robots in such a complex environment has become a research hotspot. The current common robot path planning methods mainly include global path planning and local path planning. In the aspect of global path planning, the traditional A* algorithm is widely used and can obtain a relatively optimal solution in a static environment. However, in the dynamic scene such as subway station, the A* algorithm has problems such as too many path turning times, too large angle, not considering real-time passenger flow factors, and fixed step length leading to insufficient efficiency, which is difficult to adapt to the changes of dynamic passenger flow. In the aspect of local path planning, the dynamic window method is a commonly used real-time obstacle avoidance algorithm, which selects the motion path by sampling and predicting the trajectory in the velocity space.

[0003] In the prior art, some studies attempt to realize dynamic adjustment of robot path by combining sensor mapping and passenger flow detection, but often only stay at the level of using static information of obstacles, and do not consider the real-time changes of personnel distribution, which cannot adjust the path in time according to the passenger flow density. In addition, during the peak period of the subway station, when the passenger flow density exceeds a certain threshold, the traditional method cannot generate an effective passenger evacuation scheme, which cannot effectively divert the crowd and cannot guarantee the continuous operation of the robot in the crowded environment, limiting its application value in the public transportation scene.

[0004] Therefore, how to provide a subway service robot real-time path planning method based on a graph neural network is a problem to be solved by those skilled in the art. SUMMARY

[0005] One object of the present application is to provide a subway service robot real-time path planning method based on a graph neural network. The present application provides a subway service robot real-time path planning method based on a graph neural network. The basic environment data is formed by environment mapping and rasterization, the environment and passenger flow fusion map is generated by combining target detection and personnel statistics, the global path search and optimization are realized by using the improved A* algorithm, and the redundant node deletion, circular arc smoothing processing and step dynamic adjustment are added in the path. Further, the improved dynamic window method is used for local path planning, the path cost is corrected in real time in combination with the passenger flow density, the evacuation scheme is generated when the passenger flow density is too high, and the multi-modal output is performed, so that the efficient and safe navigation of the robot in the complex passenger flow environment and the passenger evacuation guidance are realized.

[0006] A real-time path planning method for a subway service robot based on a graph neural network according to an embodiment of the present invention includes the following steps:

[0007] The robot uses its built-in lidar, ultrasonic radar, and visual camera to scan the environment inside the subway station and complete map construction. The map is then rasterized into 5cm×5cm grids, and the center point of each grid is used as the search node to obtain basic environmental data.

[0008] By using different cameras within the station to detect targets and count people, the number and distribution of passengers are obtained. Based on the camera positions and angles, the passenger distribution data is projected onto the map. Data from multiple cameras is fused and deduplicated, and then overlaid onto the basic environmental data to obtain a fused map of the environment and passenger flow.

[0009] A graph neural network model is constructed using an integrated map of the environment and passenger flow as input. The robot's starting point and target position are set, the open list and close list are initialized, the starting point node is added to the open list, obstacle nodes and high-density passenger flow nodes are added to the close list, and the initial state of global path search is output.

[0010] In the initial state of the global path search, the improved A* algorithm is called to expand the nodes, calculate the cost estimation function including path length, movement time, turning angle and safety distance, add the expanded safe nodes to the open list, and update the cost function in real time with the surrounding passenger flow density. The node with the lowest cost is selected to be added to the close list, and a candidate path from the starting point to the target point is generated.

[0011] Redundant nodes are removed from candidate paths, nodes with excessively large turning angles are smoothed with circular arcs, and the step size is dynamically adjusted according to the distribution of obstacles and passenger flow, and the optimized global path is output.

[0012] Using the optimized global path as a reference, and combining the robot's current position, speed, and direction as input, an improved dynamic window method is used for local path planning. Based on the evaluation of orientation deviation and speed, a safety distance sub-function is added to output the local path that is updated in real time.

[0013] When the passenger flow density exceeds the set threshold, a passenger evacuation plan is generated by combining the station map, passenger flow distribution and entrance / exit conditions. The plan is then output through the robot's display screen and voice broadcast to guide the evacuation of passengers and ensure the robot's passage.

[0014] Optionally, the process of obtaining the basic environmental data specifically includes:

[0015] The robot's built-in image acquisition equipment, including LiDAR, ultrasonic radar, and vision camera, is used to scan the subway station environment synchronously. The original measurement data, timestamps, and equipment calibration parameters are read. The measurement data of the three types of sensors are unified under the same reference system. The continuous measurement is registered using a synchronous positioning and mapping process. An initial map of the station containing the outline of the station structure, the position of fixed obstacles, and the robot's pose sequence is output.

[0016] The initial station map is regularly divided into units of 5 cm by 5 cm, and a grid is established with each unit uniquely identified by a row and column index. The geometric center of each grid is used as the representative position of the current grid. Based on the LiDAR echo, ultrasonic ranging return, and visual obstacle recognition results, the occupancy status of each grid is cumulatively determined. When a grid is observed as an obstacle multiple times and the observation confidence reaches the set occupancy threshold, the current grid is assigned a value of one, indicating an obstacle and impassable. When a grid is observed as free multiple times and the observation confidence reaches the set free threshold, the current grid is assigned a value of zero. When there are insufficient grid observations or the observation results do not meet the above two threshold requirements, the current grid is assigned a value of negative one. The output is a discretized grid map composed of the assigned values, as well as map information including the corresponding unit size, origin position, and index range.

[0017] The geometric center of each grid cell in the discretized raster map is used as the search node to generate a node set. The accessibility attribute of each node is recorded. Nodes with a value of zero are marked as accessible nodes, nodes with a value of one are marked as obstacle nodes, and nodes with a value of negative one are marked as unknown nodes. Adjacency relationships are established between adjacent grid centers in a four-adjacency or eight-adjacency manner to form a search topology containing the node set and edge set. The search topology, the discretized raster map, and map information are packaged together as basic environmental data.

[0018] Optionally, the process of obtaining the integrated environment and passenger flow map specifically includes:

[0019] Real-time video data is obtained by using cameras at different locations in the subway station. Pedestrian targets are detected in each video frame to identify people in the scene. A people counting algorithm is used to count the number of passengers at each camera at the corresponding time to obtain passenger flow information for different cameras.

[0020] Based on the camera's installation location and shooting angle, the detected passenger positions in the image are mapped to the station map coordinate system to form passenger position data represented by coordinates, which are then mapped to a unique raster index in the basic environmental data to obtain passenger flow distribution data indexed by the raster.

[0021] Passenger flow distribution data from different cameras are fused. Within the overlapping area of ​​the camera's field of view, passenger records that are spatially adjacent and time-close are merged to remove duplicate recognition. The number of passengers in each grid is counted on a discrete grid map according to the grid index to form a passenger flow density matrix. This matrix is ​​then overlaid with the basic environmental data to obtain an environment and passenger flow fusion map that includes obstacle information, passable areas, unknown areas, and passenger flow distribution information.

[0022] Optionally, the process of outputting the initial state of the global path search specifically includes:

[0023] A graph neural network model is constructed using an integrated environmental and passenger flow map as input. The center point of each grid in the discretized grid map is defined as a node of the graph, and the connectivity between adjacent grids is defined as an edge of the graph. The passenger flow density, obstacle occupancy, and geometric location information of each node are used as node features, and the relative orientation and travel distance of the edges are used as edge features. The trained graph neural network outputs node weights and edge weights as reference parameters for global path search.

[0024] Based on the task requirements, the robot's starting and target positions are set, and each position is located to a unique grid node in the integrated environmental and passenger flow map. The starting node and target node are generated, and a planning node set is established.

[0025] Clear the openlist and closelist. The openlist is a container for nodes to be expanded, which stores nodes to be explored. The closelist is a container for nodes that have been expanded, which stores nodes that have been fully explored.

[0026] Add the starting node to the openlist. At the same time, mark the obstacle nodes with a value of one and the passenger flow occupancy nodes with a value of one hundred in the environment and passenger flow fusion map as impassable nodes and store them in the closelist. Output the initial state of the global path search, which includes the starting node, the target node, the openlist, and the closelist.

[0027] Optionally, the generation of the candidate path specifically includes:

[0028] Taking the initial state of global path search as input, the starting node and target node, openlist, and closelist are read as the initial data for improving A* search. At the same time, obstacle nodes and passenger flow occupancy nodes are removed, and candidate nodes that are not removed are saved to openlist.

[0029] In each round of A* search, a cost estimation function is calculated for each node to be evaluated. The actual cost consists of the cumulative path length from the starting point to the current node, the cumulative movement time, and the cumulative turning angle. It is combined with the robot's safe radius and the minimum distance from the current node to the nearest obstacle to form a safety cost term. At the same time, the node weights output by the graph neural network model are called to correct the passenger flow density data. The estimated cost is given by the geometric distance between the current node and the target node. It is adjusted by combining the adaptive parameter factor and the number of nodes traversed along the coordinate axis. The actual cost and the estimated cost are added together to obtain the cost estimation function.

[0030] Select the node corresponding to the minimum cost estimation function and add it to the closelist. Record the population density of the nearby grid. If the current extended node is the same as the target node, terminate the search process and start from the target node and backtrack to the starting node step by step according to the parent node record to generate a candidate path from the starting point to the target node. If the current extended node is not the target node, continue to execute the A* search process until a complete candidate path is generated.

[0031] Optionally, the output process of the optimized global path specifically includes:

[0032] Extract nodes sequentially from the candidate path, determine whether a node is redundant. If the distance between non-adjacent nodes is less than the planned node spacing, and the straight path formed by non-adjacent nodes does not intersect with obstacles, and the perpendicular distance between the current straight path and the nearest obstacle is greater than or equal to the robot's safe radius, then the intermediate node is determined to be a redundant node, and the redundant node is deleted from the path, resulting in a sequence of retained nodes that saves the initial node, intermediate inflection point, and target node.

[0033] The corner detection is performed on the preserved node sequence. If the turning angle between three adjacent nodes is less than or equal to 90 degrees, it is determined that the current turning angle is too large. In this case, it is replaced with an arc, so that the robot transitions from a broken line to an arc curve when turning on the path, thereby reducing the turning angle and motor loss, and satisfying the robot's continuous motion constraint to generate a smooth path.

[0034] Based on the smooth path, the step size between nodes is dynamically adjusted according to the distribution of obstacles and passenger flow density. When there are many obstacles around the path or the passenger flow density is high, the step size is reduced to ensure safety. When there are few obstacles around the path and the passenger flow density is low, the step size is increased to improve movement efficiency. During the adjustment process, the passability and safety of the path are maintained. The optimized global path is output after redundant node removal, arc smoothing and dynamic step size adjustment.

[0035] Optionally, the process of outputting the real-time updated local path specifically includes:

[0036] Based on the optimized global path, the initial position, initial velocity, and attitude parameters of the robot on the global path are set, and the range of linear velocity, angular velocity, and acceleration upper limit are limited. The current position, motion velocity, and direction of the robot are sampled to establish the kinematic constraints for local path planning.

[0037] Within the velocity space, candidate linear velocities and angular velocities are sampled. For each set of candidate velocities, the robot's trajectory within a fixed time window is predicted, generating different local candidate trajectories. For each trajectory, an evaluation function for an improved dynamic window local path planning algorithm is calculated to determine the robot's forward direction under the current trajectory and compare it with the target position direction to obtain the azimuth deflection angle. The current angle difference is used as the azimuth deflection angle cost. Simultaneously, the straight-line geometric distance from the robot's current position to the target position is calculated and used as the distance cost. The linear and angular velocities corresponding to the candidate trajectories are read as velocity costs. The shortest distance between the candidate trajectory and globally known obstacles, unknown dynamic obstacles, and static obstacles is obtained. When the shortest distance is less than twice the robot's safe radius, a safe distance cost penalty is added to the current trajectory. Different grid areas with a range of four times the robot's safe radius are selected in the forward direction of the candidate trajectory. Personnel statistics are performed on each grid and personnel density is calculated. The data are superimposed to form a passenger flow density cost and assigned corresponding weight coefficients to obtain the evaluation function result.

[0038] The evaluation function results of all candidate trajectories are comprehensively evaluated, and the trajectory with the best comprehensive evaluation value is selected as the current local path output. The local path planning results are also updated cyclically in combination with the passenger flow distribution data collected in real time by the camera to maintain the dynamism and real-time nature of the path.

[0039] Optionally, the process of generating the passenger evacuation plan specifically includes:

[0040] The local path is updated in real time and combined with the environment and passenger flow fusion map to determine whether there is a usable path. If there is, collision detection continues. If there is no path, it is determined that the current passenger flow density in the station is too high, and the passenger evacuation plan generation process is initiated.

[0041] Based on the robot's current position, speed, direction of movement, and obstacle distribution information, it is determined whether a collision will occur. If the determination result is a possible collision, the obstacle attributes are further detected. If it is a dynamic obstacle, the robot returns to the local path planning step to resample the speed and predict the trajectory. If it is a static obstacle, the robot is controlled to decelerate and returns to the trajectory evaluation step to recalculate the local path. If the determination result is no risk of collision, the robot moves along the current local path.

[0042] When the number of grid cells with no available paths or passenger flow density exceeding the threshold accounts for more than 70% of the total number of grid cells in the station map, the environment and passenger flow fusion map is invoked. Combining the subway station structure layout, real-time passenger flow distribution, and entrance and exit locations, the passage capacity and path smoothness of each entrance and exit are calculated to generate a passenger evacuation plan. The path cost is also comprehensively calculated, including path length, evacuation time, personnel density, and safety distance. When the personnel density or obstacle distance does not meet the conditions, the current path cost is increased through a safety penalty function to prioritize the selection of balanced and safe evacuation paths.

[0043] The generated passenger evacuation plan is output through the robot, including displaying text and image information on the screen and playing voice prompts through the speaker. At the same time, the robot continues to perform local path planning while guiding passengers to evacuate until it reaches the target location. When it detects that the robot has reached the target location, it stops moving; otherwise, it returns to the local path planning step and continues to execute the loop.

[0044] The beneficial effects of this invention are:

[0045] 1. This invention optimizes candidate paths by introducing an improved A* algorithm in the global path planning stage. It not only comprehensively considers factors such as path length, time consumption, turning angle and safety distance on the basis of the traditional cost function, but also effectively reduces the number of turns and angles in robot motion through redundant node deletion and arc smoothing, thereby reducing energy consumption and improving motion stability.

[0046] 2. This invention adds an adaptive step length adjustment mechanism during the path planning process, which can automatically adjust the movement step length according to changes in obstacles and passenger flow distribution, so that the robot can ensure safety and improve traffic efficiency in complex environments.

[0047] 3. In the local path planning stage, this invention adopts an improved dynamic window method, which introduces the passenger flow density coefficient and the safety distance penalty function into the evaluation function, so that the path selection can respond in real time to the dynamic changes in the distribution of people in the subway station, significantly improving the obstacle avoidance capability in high-density crowd environments.

[0048] 4. When the passenger flow density in the station exceeds the threshold, the present invention can generate an evacuation plan by combining the station map, passenger flow distribution and entrance and exit conditions, and output it through a multimodal method of display and voice broadcast. This not only improves the navigation stability and continuity of the robot in the case of large passenger flow, but also realizes the effective guidance of passengers and improves the safety and efficiency of the overall station operation. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Fig. 1 This is a flowchart of a real-time path planning method for a subway service robot based on a graph neural network proposed in this invention.

[0051] Fig. 2 This is a schematic diagram of global path planning for a real-time path planning method for subway service robots based on graph neural networks proposed in this invention.

[0052] Fig. 3 This is a schematic diagram of local path planning for a real-time path planning method for subway service robots based on graph neural networks proposed in this invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figs. 1-3 A real-time path planning method for subway service robots based on graph neural networks includes the following steps:

[0055] The robot uses its built-in lidar, ultrasonic radar, and visual camera to scan the environment inside the subway station and complete map construction. The map is then rasterized into 5cm×5cm grids, and the center point of each grid is used as the search node to obtain basic environmental data.

[0056] By using different cameras within the station to detect targets and count people, the number and distribution of passengers are obtained. Based on the camera positions and angles, the passenger distribution data is projected onto the map. Data from multiple cameras is fused and deduplicated, and then overlaid onto the basic environmental data to obtain a fused map of the environment and passenger flow.

[0057] A graph neural network model is constructed using an integrated map of the environment and passenger flow as input. The robot's starting point and target position are set, the open list and close list are initialized, the starting point node is added to the open list, obstacle nodes and high-density passenger flow nodes are added to the close list, and the initial state of global path search is output.

[0058] In the initial state of the global path search, the improved A* algorithm is called to expand the nodes, calculate the cost estimation function including path length, movement time, turning angle and safety distance, add the expanded safe nodes to the open list, and update the cost function in real time with the surrounding passenger flow density. The node with the lowest cost is selected to be added to the close list, and a candidate path from the starting point to the target point is generated.

[0059] Redundant nodes are removed from candidate paths, retaining only the starting point, key turning points, and target point. Nodes with excessively large turning angles are smoothed with circular arcs, and the step size is dynamically adjusted according to obstacles and passenger flow distribution to output the optimized global path.

[0060] Using the optimized global path as a reference, and combining the robot's current position, speed, and direction as input, an improved dynamic window method is used for local path planning. Based on the evaluation of orientation deviation and speed, a safety distance sub-function is added to output the local path that is updated in real time.

[0061] When the passenger flow density exceeds the set threshold, a passenger evacuation plan is generated by combining the station map, passenger flow distribution and entrance / exit conditions. The plan is then output through the robot's display screen and voice broadcast to guide the evacuation of passengers and ensure the robot's passage.

[0062] This invention combines data from lidar, ultrasonic radar, and visual cameras to achieve accurate mapping of the subway station environment. It dynamically optimizes path planning using passenger flow density information, and improves the A* algorithm by combining it with the dynamic window method. This allows the robot to adjust the path according to real-time passenger flow density, ensuring safe and efficient operation in complex environments. Redundant node removal, arc smoothing, and adaptive step size adjustment significantly improve the smoothness of the path and the efficiency of movement. When the passenger flow density exceeds a threshold, it automatically generates and outputs a passenger evacuation plan, ensuring both passenger safety and smooth robot passage, thus enhancing the adaptability and reliability of robot navigation within subway stations.

[0063] In this embodiment, the process of obtaining the basic environmental data specifically includes:

[0064] The robot's built-in image acquisition equipment, including LiDAR, ultrasonic radar, and vision camera, is used to scan the subway station environment synchronously. The original measurement data, timestamps, and equipment calibration parameters are read. The measurement data of the three types of sensors are unified under the same reference system. The continuous measurement is registered using a synchronous positioning and mapping process. An initial map of the station containing the outline of the station structure, the position of fixed obstacles, and the robot's pose sequence is output.

[0065] The initial station map is regularly divided into units of 5 cm by 5 cm, and a grid is established with each unit uniquely identified by a row and column index. The geometric center of each grid is used as the representative position of the current grid. Based on the LiDAR echo, ultrasonic ranging return, and visual obstacle recognition results, the occupancy status of each grid is cumulatively determined. When the current grid is observed as an obstacle multiple times and the observation confidence reaches the set occupancy threshold, the current grid is assigned a value of one, indicating an obstacle and impassable. When the grid is observed as free multiple times and the observation confidence reaches the set free threshold, the current grid is assigned a value of zero. When the grid observation is insufficient or the observation results do not meet the above two threshold requirements, the current grid is assigned a value of negative one. The output is a discretized grid map composed of the assigned values, as well as map information including the corresponding unit size, origin position, and index range.

[0066] Using the geometric center of each grid cell in the discretized raster map as the search node, a node set is generated, and a passage attribute is recorded for each node. Nodes with a value of zero are marked as passable nodes, nodes with a value of one are marked as obstacle nodes, and nodes with a value of negative one are marked as unknown nodes. Adjacency relationships are established between adjacent grid centers in a four-adjacency or eight-adjacency manner to form a search topology containing a node set and an edge set. This topology, along with the search topology, the discretized raster map, and map information, is packaged together as basic environmental data.

[0067] This invention achieves high-precision synchronous scanning and mapping of the subway station environment by combining data from lidar, ultrasonic radar, and visual cameras. It uses a synchronous positioning and mapping process to register sensor data, generate an initial map of the station, and discretize the environment through regularized rasterization. Through multiple observations and confidence level determinations, it accurately identifies obstacles and vacant areas, providing reliable environmental information for path planning. The raster map generated by this method not only improves the accuracy of environmental mapping but also updates the dynamic changes of obstacles and passage areas in real time, optimizing the robot's navigation capabilities in complex environments.

[0068] In this embodiment, the process of obtaining the integrated environment and passenger flow map specifically includes:

[0069] Real-time video data is obtained by using cameras at different locations in the subway station. Pedestrian targets are detected in each video frame to identify people in the scene. A people counting algorithm is used to count the number of passengers at each camera at the corresponding time to obtain passenger flow information for different cameras.

[0070] Based on the camera's installation location and shooting angle, the detected passenger positions in the image are mapped to the station map coordinate system to form passenger position data represented by coordinates, which are then mapped to a unique raster index in the basic environmental data to obtain passenger flow distribution data indexed by the raster.

[0071] Passenger flow distribution data from different cameras are fused. Within the overlapping area of ​​the camera's field of view, passenger records that are spatially adjacent and time-close are merged to remove duplicate recognition. The number of passengers in each grid is counted on a discrete grid map according to the grid index to form a passenger flow density matrix. This matrix is ​​then overlaid with the basic environmental data to obtain an environment and passenger flow fusion map that includes obstacle information, passable areas, unknown areas, and passenger flow distribution information.

[0072] This invention combines real-time video data from different cameras within a subway station, utilizing pedestrian target detection and people counting algorithms to accurately obtain the number and distribution of passengers in each area. By mapping passenger locations to the station's map coordinate system and fusing them with basic environmental data, an environmental map containing passenger flow information is formed. Employing multi-camera data fusion technology eliminates duplicate identification and effectively counts the passenger density of each grid cell, generating a passenger flow density matrix.

[0073] In this embodiment, the process of outputting the initial state of the global path search specifically includes:

[0074] A graph neural network model is constructed using an integrated environmental and passenger flow map as input. The center point of each grid in the discretized grid map is defined as a node of the graph, and the connectivity between adjacent grids is defined as an edge of the graph. The passenger flow density, obstacle occupancy, and geometric location information of each node are used as node features, and the relative orientation and travel distance of the edges are used as edge features. The trained graph neural network outputs node weights and edge weights as reference parameters for global path search.

[0075] Based on the task requirements, the robot's starting and target positions are set, and each position is located to a unique grid node in the integrated environmental and passenger flow map. The starting node and target node are generated, and a planning node set is established.

[0076] Clear the openlist and closelist. The openlist is a container for nodes to be expanded, which stores nodes to be explored. The closelist is a container for nodes that have been expanded, which stores nodes that have been fully explored.

[0077] Add the starting node to the openlist. At the same time, mark the obstacle nodes with a value of one and the passenger flow occupancy nodes with a value of one hundred in the environment and passenger flow fusion map as impassable nodes and store them in the closelist. Output the initial state of the global path search, which includes the starting node, the target node, the openlist, and the closelist.

[0078] This invention constructs a graph neural network model based on an environment and passenger flow fusion map, inputs node and edge features into the network and outputs weight results, providing dynamic reference parameters for global path search. It can simultaneously consider topological relationships, obstacle distribution and passenger flow density in complex subway scenarios, and realize intelligent modeling of the path search space. It effectively overcomes the shortcomings of traditional algorithms in handling dynamic passenger flow and complex topological structures, making the generated paths more reasonable and smoother, and with stronger adaptability and robustness.

[0079] In this embodiment, the generation of the candidate path specifically includes:

[0080] Taking the initial state of global path search as input, the starting node and target node, openlist, and closelist are read as the initial data for improving A* search. At the same time, obstacle nodes and passenger flow occupancy nodes are removed, and candidate nodes that are not removed are saved to openlist.

[0081] In each round of A* search, a cost estimation function is computed for each node to be evaluated:

[0082] F(n) = G(n) + H(n);

[0083]

[0084]

[0085] Where F(n) is the cost estimation function used to estimate the cost of reaching the target state from the initial state through state n, G(n) is the actual cost from the starting point to node n in the state space, H(n) is the estimated cost of the optimal path from node n to the target state, d(i) is the actual cost distance from the starting node to the current node i, t(i) is the actual time taken from the starting node to the current node, a(i) is the actual turning angle from the starting node to the current node, is the robot's safe radius, min(di) is the minimum distance from the current node to the nearest obstacle on the path between the current node and the next node, μ is the safety cost weight, max is the maximum value function, k is the adaptive parameter factor, and M... x and M y Here, x represents the number of nodes traversed by the mobile robot along the x-axis and y-axis, respectively. max x min y max y min P represents the maximum and minimum nodes along the x-axis and y-axis, respectively. ij This is a binary indicator variable used to determine whether the position (i,j) is an obstacle, 1-P ijh1(n) represents the number of unobstructed nodes between the starting node and the target node, and h1(n) represents the distance between the current node and the target node. The actual cost consists of the cumulative path length from the starting point to the current node, the cumulative movement time, and the cumulative turning angle. It is combined with the robot's safe radius and the minimum distance from the current node to the nearest obstacle to form a safety cost term. At the same time, the node weights output by the graph neural network model are called to correct the passenger flow density data. The estimated cost is given by the geometric distance between the current node and the target node, and is adjusted by combining the adaptive parameter factor and the number of nodes traversed along the coordinate axis. The actual cost and the estimated cost are added together to obtain the cost estimation function.

[0086] Select the node corresponding to the minimum cost estimation function and add it to the closelist. Record the population density of the nearby grid. If the current extended node is the same as the target node, terminate the search process and start from the target node and backtrack to the starting node step by step according to the parent node record to generate a candidate path from the starting point to the target node. If the current extended node is not the target node, continue to execute the A* search process until a complete candidate path is generated.

[0087] This invention improves the A* algorithm by comprehensively considering factors such as the robot's actual cost, turning angle, safety distance, and passenger density during path search, effectively optimizing path planning. The introduction of safety cost weights and passenger density corrections enhances the robot's obstacle avoidance capabilities in dynamic environments and ensures path accessibility and safety. By adjusting path planning with adaptive parameters, the invention improves the robot's real-time response capabilities in complex subway stations, significantly enhancing the efficiency and stability of path planning and ensuring efficient navigation of the robot in high-passenger-flow scenarios.

[0088] In this embodiment, the output process of the optimized global path specifically includes:

[0089] Extract nodes sequentially from the candidate path, determine whether a node is redundant. If the distance between non-adjacent nodes is less than the planned node spacing, and the straight path formed by non-adjacent nodes does not intersect with obstacles, and the perpendicular distance between the current straight path and the nearest obstacle is greater than or equal to the robot's safe radius, then the intermediate node is determined to be a redundant node, and the redundant node is deleted from the path, resulting in a sequence of retained nodes that saves the initial node, intermediate inflection point, and target node.

[0090] The corner detection is performed on the preserved node sequence. If the turning angle between three adjacent nodes is less than or equal to 90 degrees, it is determined that the current turning angle is too large. In this case, it is replaced with an arc, so that the robot transitions from a broken line to an arc curve when turning on the path, thereby reducing the turning angle and motor loss, and satisfying the robot's continuous motion constraint to generate a smooth path.

[0091] Based on the smooth path, the step size between nodes is dynamically adjusted according to the distribution of obstacles and passenger flow density. When there are many obstacles around the path or the passenger flow density is high, the step size is reduced to ensure safety. When there are few obstacles around the path and the passenger flow density is low, the step size is increased to improve movement efficiency. During the adjustment process, the passability and safety of the path are maintained. The optimized global path is output after redundant node removal, arc smoothing and dynamic step size adjustment.

[0092] This invention achieves a simpler and smoother path structure through redundant node removal, arc smoothing, and dynamic step size adjustment during global path optimization. This reduces unnecessary turns and stops for the robot, thereby reducing path complexity and motor energy consumption while improving the continuity and stability of robot operation. Furthermore, by combining obstacle distribution and passenger flow density for adaptive step size adjustment, it not only ensures safety in high-density environments but also enhances operational efficiency in sparse areas, thus achieving efficient and reliable global path navigation.

[0093] In this embodiment, the process of outputting the real-time updated local path specifically includes:

[0094] Based on the optimized global path, the initial position, initial velocity, and attitude parameters of the robot on the global path are set, and the range of linear velocity, angular velocity, and acceleration upper limit are limited. The current position, motion velocity, and direction of the robot are sampled to establish the kinematic constraints for local path planning.

[0095] The candidate linear and angular velocities are sampled in the velocity space. For each set of candidate velocities, the robot's trajectory is predicted within a fixed time window in the future, generating different local candidate trajectories. For each trajectory, an evaluation function for the improved dynamic window local path planning algorithm is calculated.

[0096] G(v,ω)=σ[α·heading(v,ω)+β·(density·Dist(v,ω)+ε·penalty)+γ·

[0097] velocity(v,ω);

[0098] Wherein, G(v,ω) is the evaluation function, v and ω are the linear velocity and angular velocity, respectively; heading(v,ω) is the orientation deflection angle evaluation sub-function, representing the angle difference between the robot's forward direction and the target point direction; density·Dist(v,ω)+ε·penalty is the safety distance evaluation sub-function, where Dist(v,ω) is the shortest distance between globally known obstacles, unknown dynamic obstacles, and static obstacles and the simulated path; density is the personnel density coefficient within 100×100 grids outside 4 times the safety radius of the robot's forward direction; ε·penalty is the safety penalty function; velocity(v,ω) is the velocity evaluation function, representing the speed of the mobile robot in its current state; and α, β, and ε are weighting coefficients used to adjust the impact of factors such as the mobile robot's motion direction, distance, and speed on the evaluation function results. The influence of σ is used to normalize the function result, making the path relatively smooth; the robot's forward direction under the current trajectory is determined and compared with the target position direction to obtain the azimuth deflection angle. The current angle difference is used as the azimuth deflection angle cost. At the same time, the straight-line geometric distance from the robot's current position to the target position is calculated and used as the distance cost. The linear velocity and angular velocity corresponding to the candidate trajectory are read as the velocity cost. The shortest distance between the candidate trajectory and globally known obstacles, unknown dynamic obstacles, and static obstacles is obtained. When the shortest distance is less than twice the robot's safe radius, a safe distance cost penalty is added to the current trajectory. Different grid areas with a range of four times the robot's safe radius are selected in the forward direction of the candidate trajectory. The personnel are counted and the personnel density is calculated for each grid. The personnel density is superimposed to form the passenger flow density cost and assigned a corresponding weight coefficient to obtain the evaluation function result.

[0099] The evaluation function results of all candidate trajectories are comprehensively evaluated, and the trajectory with the best comprehensive evaluation value is selected as the current local path output. The local path planning results are also updated cyclically in combination with the passenger flow distribution data collected in real time by the camera to maintain the dynamism and real-time nature of the path.

[0100] This invention introduces an improved dynamic window algorithm into local path planning, comprehensively considering multiple factors such as azimuth deflection angle, target distance, speed, safety distance, and passenger flow density in trajectory evaluation, and combining a safety penalty mechanism with personnel density correction, thereby achieving stable obstacle avoidance and path optimization in high-density crowd and dynamic obstacle environments.

[0101] In this embodiment, the process of generating the passenger evacuation plan specifically includes:

[0102] The local path is updated in real time and combined with the environment and passenger flow fusion map to determine whether there is a usable path. If there is, collision detection continues. If there is no path, it is determined that the current passenger flow density in the station is too high, and the passenger evacuation plan generation process is initiated.

[0103] Based on the robot's current position, speed, direction of movement, and obstacle distribution information, it is determined whether a collision will occur. If the determination result is a possible collision, the obstacle attributes are further detected. If it is a dynamic obstacle, the robot returns to the local path planning step to resample the speed and predict the trajectory. If it is a static obstacle, the robot is controlled to decelerate and returns to the trajectory evaluation step to recalculate the local path. If the determination result is no risk of collision, the robot moves along the current local path.

[0104] When the number of grid cells with no available paths or passenger flow density exceeding the threshold accounts for more than 70% of the total number of grid cells in the station map, the environment and passenger flow fusion map is invoked. Combining the subway station structure layout, real-time passenger flow distribution, and entrance and exit locations, the passage capacity and path smoothness of each entrance and exit are calculated to generate a passenger evacuation plan. The path cost is also comprehensively calculated, including path length, evacuation time, personnel density, and safety distance. When the personnel density or obstacle distance does not meet the conditions, the current path cost is increased through a safety penalty function to prioritize the selection of balanced and safe evacuation paths.

[0105] The generated passenger evacuation plan is output through the robot, including displaying text and image information on the screen and playing voice prompts through the speaker. At the same time, the robot continues to perform local path planning while guiding passengers to evacuate until it reaches the target location. When it detects that the robot has reached the target location, it stops moving; otherwise, it returns to the local path planning step and continues to execute the loop.

[0106] This invention enables efficient path planning for service robots in complex and dynamic environments such as subway stations. By integrating environmental maps and passenger flow distribution information for real-time path updates, it generates passenger evacuation plans and outputs them in a multimodal manner when feasible paths are insufficient or passenger flow density is too high, ensuring dual safety for passenger guidance and robot passage. The improved collision detection and dynamic window planning methods effectively enhance the robot's obstacle avoidance capabilities and navigation stability in high-density passenger flow scenarios, thereby enhancing the overall adaptability, reliability, and practicality of path planning.

[0107] Example 1:

[0108] To verify the feasibility of this invention in practice, it was applied to a subway station service robot guiding passengers to the entrance and exit during peak hours. The aim is to solve the problems of slow response, unreasonable path planning, and robots easily getting stuck in congested areas in dynamic crowd environments by traditional path planning algorithms. This embodiment demonstrates the improved efficiency and safety of path planning in complex environments by comparing experimental data.

[0109] During morning and evening rush hours, subway stations experience a significant increase in passenger flow, especially in station halls, transfer passages, and entrance / exit areas, where passenger density is high and flow directions are complex. Service robots need to complete tasks such as information guidance and assisted evacuation without interfering with the normal passage of passengers. Previously used path planning algorithms often did not take into account real-time passenger flow information, causing robots to traverse high-density crowd areas, resulting in frequent obstacle avoidance, pauses, or even standing still, which greatly reduced service efficiency.

[0110] In this scenario, the invention combines four core modules: SLAM map building, improved global path planning (A), improved local path adjustment (DWA), and station passenger flow recognition. The service robot first scans and maps the station environment using its onboard LiDAR and visual sensors, then rasterizes the map, dividing it into 5cm x 5cm grids marked with accessible, obstacle, and unknown areas. Simultaneously, multiple cameras deployed in the subway station transmit captured images to a backend server. Object detection and crowd counting algorithms are used to obtain real-time passenger flow density in various areas of the station. This data is overlaid with the raster map to form a dynamic passenger flow map. Upon receiving a task instruction, the robot activates the improved A global path planning algorithm. In the initial path planning, passenger flow density and grid safety weights are considered, prioritizing avoidance of crowded areas. During path execution, the robot samples in real-time based on its current location and dynamically corrects the local path using the improved DWA algorithm to avoid newly appearing crowds or obstacles. If the passenger flow density within the station reaches a warning threshold, the system automatically triggers a passenger evacuation algorithm. The robot will broadcast guidance voice and output suggested routes along the evacuation direction.

[0111] Table 1. Performance Comparison of the Invention Method and Traditional Path Planning Methods

[0112]

[0113] This invention avoids congested areas more effectively in path planning. Although the path is slightly longer, the actual task completion time is significantly reduced, reflecting a more efficient overall path. The number of pauses is reduced by 75%, indicating that this invention effectively avoids the phenomenon of frequent stops caused by the robot's inability to avoid congested environments. The frequency of close contact with passengers is greatly reduced, verifying the effectiveness of the safe distance model and density penalty function. The total turning angle and smoothness of the path are significantly optimized, improving the continuity and stability of the robot's movement and reducing motor losses. The dynamic obstacle avoidance and real-time adjustment strategy enables the robot to maintain near-ideal speed in complex environments, enhancing its ability to adapt to environmental changes.

[0114] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A real-time path planning method for subway service robots based on graph neural networks, characterized in that, Includes the following steps: The robot's built-in image acquisition device is used to scan the environment inside the subway station to complete map construction. The map is then rasterized to obtain basic environmental data. By using different cameras within the station to detect targets and count people, the number and distribution of passengers are obtained. Data from multiple cameras is fused and deduplicated, and then overlaid onto the basic environmental data to obtain a fused map of the environment and passenger flow. A graph neural network model is constructed using an integrated map of the environment and passenger flow as input. The robot's starting point and target position are set, the open list and close list are initialized, and the initial state of the global path search is output. In the initial state of the global path search, the improved A* algorithm is invoked to expand nodes, calculate the cost estimation function, and generate candidate paths; Redundant nodes are removed from candidate paths, nodes with excessively large turning angles are smoothed with circular arcs, and the step size is dynamically adjusted according to the distribution of obstacles and passenger flow, and the optimized global path is output. An improved dynamic window method is used to plan local paths for the optimized global path, and the local paths are updated in real time. By combining the station map, passenger flow distribution, and entrance / exit conditions, a passenger evacuation plan is generated and output through the robot's display screen and voice broadcast to guide passenger evacuation and ensure the robot's passage.

2. The real-time path planning method for a subway service robot based on a graph neural network according to claim 1, characterized in that, The process of obtaining the basic environmental data specifically includes: The robot's built-in image acquisition device is used to scan the subway station environment synchronously. The synchronous positioning and mapping process is used to register the continuous measurements and output the initial station map. The initial site map is divided into regular sections, and a grid is established with each grid cell uniquely identified by its row and column indices. The geometric center of each grid cell is used as the representative position of the current grid cell. The occupancy status of each grid cell is cumulatively determined, and a discretized grid map and map information are output. Using the geometric center of each grid in the discretized raster map as the search node, a node set is generated, and adjacency relationships are established for adjacent grid centers to form a search topology. This search topology, along with the discretized raster map and map information, is packaged together as basic environmental data.

3. The real-time path planning method for a subway service robot based on a graph neural network according to claim 1, characterized in that, The process of obtaining the integrated environment and passenger flow map specifically includes: Real-time video data is obtained by using cameras at different locations in the subway station. Pedestrian targets are detected in each video frame, and the number of passengers at each camera at the corresponding time is counted to obtain passenger flow information. Based on the camera's installation location and shooting angle, the detected passenger positions in the image are mapped to the station map coordinate system to form passenger position data, which is then mapped to a unique raster index in the basic environmental data to obtain passenger flow distribution data. Passenger flow distribution data from different cameras are fused and processed, and the number of passengers in each grid is counted according to the grid index on the discretized grid map to form a passenger flow density matrix. This matrix is ​​then overlaid with the basic environmental data to obtain an environment and passenger flow fusion map.

4. The real-time path planning method for a subway service robot based on a graph neural network according to claim 1, characterized in that, The process of outputting the initial state of the global path search specifically includes: A graph neural network model is constructed using an integrated environmental and passenger flow map as input. The center point of each grid in the discretized grid map is defined as a node of the graph, and the connectivity between adjacent grids is defined as an edge of the graph. The passenger flow density, obstacle occupancy, and geometric location information of each node are used as node features, and the relative orientation and travel distance of the edges are used as edge features. The trained graph neural network outputs node weights and edge weights as reference parameters for global path search. Based on the task requirements, the robot's starting and target positions are set, and each position is located to a unique grid node in the integrated environmental and passenger flow map. The starting node and target node are generated, and a planning node set is established. Clear the openlist and closelist. The openlist is a container for nodes to be expanded, which stores nodes to be explored. The closelist is a container for nodes that have been expanded, which stores nodes that have been fully explored. Add the starting node to the openlist, and mark the obstacle nodes with a value of one and the passenger flow occupancy nodes with a value of one hundred in the environment and passenger flow fusion map as impassable nodes, store them in the closelist, and output the initial state of the global path search.

5. The real-time path planning method for a subway service robot based on a graph neural network according to claim 1, characterized in that, The generation of the candidate path specifically includes: Taking the initial state of global path search as input, the starting node, target node, openlist, and closelist are read as the initial data for improving A* search. At the same time, obstacle nodes and passenger flow occupancy nodes are removed, and candidate nodes that are not removed are saved to openlist. In each round of A* search, a cost estimation function is calculated for each node to be evaluated. The actual cost consists of the cumulative path length from the starting point to the current node, the cumulative movement time, and the cumulative turning angle. It is combined with the robot's safe radius and the minimum distance from the current node to the nearest obstacle to form a safety cost term. At the same time, the node weights output by the graph neural network model are called to correct the passenger flow density data. The estimated cost is given by the geometric distance between the current node and the target node. It is adjusted by combining the adaptive parameter factor and the number of nodes traversed along the coordinate axis. The actual cost and the estimated cost are added together to obtain the cost estimation function. Select the node corresponding to the minimum cost estimation function and add it to the closelist. Record the population density of the nearby grid. If the current expanded node is the same as the target node, terminate the search process and start from the target node to backtrack to the starting node step by step according to the parent node record to generate a candidate path from the starting node to the target node.

6. The real-time path planning method for a subway service robot based on a graph neural network according to claim 1, characterized in that, The optimized global path output process specifically includes: Extract nodes sequentially from the candidate path, determine whether a node is redundant, delete redundant nodes from the path, and obtain a sequence of retained nodes that preserves the initial node, intermediate inflection points and target node. The reserved node sequence is subjected to corner detection. If the turning angle between three adjacent nodes is too large, it is replaced with an arc to generate a smooth path. Based on the smooth path, the step size between nodes is dynamically adjusted according to the distribution of obstacles and passenger flow density, and the optimized global path is output.

7. The real-time path planning method for a subway service robot based on a graph neural network according to claim 1, characterized in that, The process of outputting the real-time updated local path specifically includes: Based on the optimized global path, the robot's current position, speed and direction are sampled to establish kinematic constraints for local path planning. Within the velocity space, candidate linear velocities and angular velocities are sampled to generate different local candidate trajectories. For each trajectory, an evaluation function of the improved dynamic window local path planning algorithm is calculated to determine the robot's forward direction under the current trajectory and compare it with the target position direction to obtain the azimuth deflection angle. At the same time, the straight-line geometric distance from the robot's current position to the target position is calculated and used as the distance cost. The linear velocity and angular velocity corresponding to the candidate trajectory are read as the velocity cost. The shortest distance between the candidate trajectory and globally known obstacles, unknown dynamic obstacles, and static obstacles is obtained. When the shortest distance is less than twice the robot's safe radius, a safe distance cost penalty is added to the current trajectory. Different grid areas with a range of four times the robot's safe radius are selected in the forward direction of the candidate trajectory. Personnel statistics are performed on each grid and personnel density is calculated. The data are superimposed to form a passenger flow density cost and assigned corresponding weight coefficients to obtain the evaluation function result. The evaluation function results of all candidate trajectories are comprehensively evaluated, and the trajectory with the best comprehensive evaluation value is selected as the current local path output.

8. The real-time path planning method for a subway service robot based on a graph neural network according to claim 1, characterized in that, The process of generating the passenger evacuation plan specifically includes: By combining real-time updated local paths with an integrated map of the environment and passenger flow, it is determined whether there are any available paths. If no path is found, it is determined that the current passenger flow density in the station is too high, and the process of generating a passenger evacuation plan is initiated. Based on the robot and obstacle distribution information, determine whether a collision will occur. If the determination result is no risk of collision, the robot moves forward along the current local path. If no available path is determined or the passenger flow density exceeds the threshold, the environment and passenger flow fusion map is invoked to calculate the passage capacity and path smoothness of each entrance and exit, and generate a passenger evacuation plan. The generated passenger evacuation plan is output through the robot. While guiding the passengers to evacuate, the robot continues to perform local path planning until it reaches the target location. When it is detected that the robot has reached the target location, it stops moving; otherwise, it returns to the local path planning step and continues to execute in a loop.

Images