Vehicle-road cooperation path planning decision-making system based on deep learning

Abstract

The invention discloses a vehicle-road cooperation path planning decision-making system based on deep learning, and belongs to the technical field of intelligent traffic. According to the system, global traffic flow data, vehicle positions and vehicle driving destinations are collected, traffic flow prediction of each road section is carried out in combination with a GNN-LSTM fusion model, and a basis is provided for global path planning; the motion trails of the participants are predicted through a Transform model, and real-time interactive decision making between the vehicles and the participants is achieved through a DQN algorithm; vehicle state data, traffic light time sequence data and intersection geographic parameters are fused based on an MPC algorithm to generate a vehicle intersection passing track, whether a dangerous space-time window exists or not is detected through an IoU algorithm, whether the MPC algorithm with constraints is adopted to correct the vehicle intersection passing track or not is determined, and intersection collaborative decision making is achieved. According to the invention, the dynamic adaptability of vehicle path planning is improved, the safety of vehicle-road cooperation and interaction with surrounding participants is enhanced, and the traffic efficiency is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation technology, and specifically relates to a vehicle-road collaborative path planning decision system based on deep learning. Background Art

[0002] With the rapid development of intelligent transportation technology, vehicle routing and vehicle-infrastructure collaboration have become core research areas for improving traffic efficiency and safety. Existing vehicle routing systems often make decisions based on static road information or local traffic data. These systems suffer from insufficient dynamic adaptability, weak vehicle-infrastructure collaboration capabilities, and low security in interactions with surrounding participants, making them difficult to meet the demands of complex traffic scenarios.

[0003] In terms of path planning, traditional systems rely on static road network models and fixed-weight algorithms, and are unable to respond to dynamic changes in global traffic flow in real time. They are prone to falling into local optimality, resulting in a decline in overall traffic efficiency. In terms of vehicle-road collaboration and participant interaction, the trajectory prediction accuracy of surrounding participants is insufficient, and there is a lack of in-depth exploration of historical laws and spatial correlations, resulting in delayed or misjudgment of driving fine-tuning decisions. At the same time, the collaboration between vehicles and road infrastructure is mostly one-way information reception, and there is a lack of dynamic trajectory collaboration based on real-time vehicle status and signal light timing, which can easily lead to "green light idling" or "red light rushing". In the intersection decision-making link, the trigger mechanism is not combined with the dynamic adjustment of the intersection type, resulting in decision-making initiated too early or too late, and trajectory planning and correction lack hard constraints on dangerous time and space windows, making it difficult to generate safe and efficient corrected trajectories in a timely manner, which can easily lead to traffic conflicts.

[0004] In summary, existing technologies have obvious shortcomings in dynamic traffic adaptation, deep vehicle-road collaboration, safe interaction between participants, and accurate decision-making at intersections. There is an urgent need for a vehicle-road collaborative path planning system that integrates deep learning technology and has global perception and collaborative decision-making capabilities to improve traffic efficiency and driving safety. Summary of the Invention

[0005] The purpose of the present invention is to provide a vehicle-road collaborative path planning decision system based on deep learning to solve the problems raised in the above background technology.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: A deep learning-based vehicle-road collaborative path planning and decision-making system, including a data perception and acquisition module, a global path planning module, a participant interaction decision-making module, an intersection scenario triggering module, a road infrastructure collaboration module, an intersection comprehensive collaborative decision-making module, and a vehicle execution control system; The data sensing and acquisition module is used to obtain vehicle status data, vehicle travel destination, participant action data, global traffic flow data, traffic light timing data, road topology and intersection geographical parameters in real time; The global path planning module is used to build a traffic network model based on road topology and global traffic flow data, and to perform a path search on the traffic network model using an improved A* algorithm in combination with the vehicle's destination and location to obtain several candidate paths. The module then performs multi-objective optimization on the candidate paths to obtain the optimal planned path. The participant interaction decision module is used to obtain driving fine-tuning action instructions based on the participant action data and vehicle status data; The intersection scenario trigger module is used to calculate the vehicle intersection distance based on the vehicle position and intersection coordinates, compare the vehicle intersection distance with a preset distance threshold, determine whether to synchronously trigger the road infrastructure collaboration module and the intersection comprehensive collaborative decision module, and push the blocking driving fine-tuning action instruction to the vehicle execution control system; The road infrastructure collaboration module is used to obtain the trajectory of vehicles passing through the intersection based on vehicle status data, traffic light timing data and intersection geographical parameters; The intersection comprehensive collaborative decision-making module is used to calculate the dangerous time-space window of the vehicle's trajectory through the intersection, and if the dangerous time-space window exists, the vehicle's trajectory through the intersection is corrected; The vehicle execution control system is used to control the vehicle to execute the optimal planned path, as well as driving fine-tuning action instructions or the vehicle's trajectory through the intersection.

[0007] Preferably, the data perception and acquisition module serves as the basic data input end, realizes comprehensive perception of traffic scene information through multi-source fusion, and obtains vehicle status data and participant motion data in real time through on-board sensors. The on-board sensors include lidar, millimeter-wave radar, camera, wheel speed sensor, acceleration sensor and steering angle sensor, among which the lidar and millimeter-wave radar are responsible for real-time collection of the three-dimensional position coordinates and relative speed of participants including pedestrians and other vehicles within a 360-degree range around the vehicle. The camera extracts the movement direction and posture characteristics of the participants through image recognition technology, and forms a continuous position time series sequence, speed time series sequence and direction time series sequence of the participants after timestamp synchronization, which constitute the participant motion data. At the same time, the vehicle status data is obtained in real time through the on-board CAN bus, among which the vehicle position is output by the GPS positioning module combined with RTK technology, the vehicle speed and vehicle acceleration are jointly measured by the wheel speed sensor and the acceleration sensor, and the vehicle direction is calculated by the fusion of the electronic compass and the steering angle sensor; V2X interacts with the roadside unit (RSU) via PC5 direct communication to obtain global traffic flow data and traffic light timing data. Global traffic flow data covers the entire traffic network, including the total number of vehicles on each road section, section length, average speed, and congestion level. Traffic light timing data includes signal light status, remaining signal light status time, and signal light timing plan for the next three cycles. The in-vehicle navigation system interface is connected to the map service platform through OTA to obtain the latitude and longitude coordinates of the vehicle's destination, road topology and intersection geographical parameters including intersection type, lane width of each lane at the intersection, intersection boundary coordinates, lane line position and intersection coordinates. The intersection coordinates are the longitude and latitude coordinates of the midpoint of the intersection stop line, which serve as the reference point for triggering intersection scenario decisions.

[0008] Preferably, the specific method for constructing a traffic network model based on road topology and global traffic flow data is: S1. The length of each road segment in the global traffic flow data , average speed and congestion levels Calculate the section weight of each section using the section weight formula ,The road topology is formed with the intersections as nodes and the sections between the intersections as edges, and the section weight of each section is used as the edge weight of the corresponding edge in the road topology, and the initial ,model of the traffic network is obtained; The road section weight formula is: ; in, is the adjustment parameter; S2. The initial traffic network model is analyzed and predicted using the GNN-LSTM fusion model. The initial traffic network model is considered as a graph structure, with nodes as intersections, edges as road sections, and edge features including the average vehicle speed of each road section. and congestion levels Based on this, the spatial correlation between road sections is extracted through the graph neural network (GNN), and the road section state vector that integrates spatial features is output. The LSTM layer inputs the road section state vector output by the GNN in time series, learns the time evolution law of the road section traffic flow, and finally outputs the predicted average speed and predicted congestion level of each road section; S3. Substitute the predicted average vehicle speed and predicted congestion level of each road section output by S2, as well as the corresponding road section length, into the road section weight formula in S1 to calculate the updated road section weight of each road section. The updated road section weight is reassigned to the edge weight of the edge corresponding to the road topology to obtain the traffic network model.

[0009] Preferably, the process of performing path search on the traffic network model by using the improved A* algorithm is as follows: Improve the heuristic function of the A* algorithm Based on the above, the traffic flow balance factor is introduced. ; The heuristic function of the improved A* algorithm is: ; in, is the comprehensive cost assessment value, is the actual cost from the starting point to the current node which is the middle point or intersection of the road segment, is the estimated cost from the current node to the end point, is the adjustment coefficient, which takes ; The improved A* algorithm uses the vehicle location as the starting point and the vehicle destination as the end point to perform a depth-first search path search on the traffic network model. When the first candidate path is found, that is, a combination of continuous road sections from the starting point to the end point, the path cost value of the first candidate path is calculated and defined as the optimal candidate path cost value. , where the path cost value of the first candidate path is the sum of the segment weights of all segments in the path; After the first candidate path search is completed, continue to search for other candidate paths, and only keep the path cost value in The paths within the range are selected as candidate paths, where is the preset cost value threshold and its value is If the number of candidate paths found is less than 3, the To 0.5.

[0010] Preferably, a method for performing multi-objective optimization on several candidate paths is as follows: Each candidate path in several candidate paths is split into sections, and each section included in each candidate path and the corresponding total number of vehicles and section weight are obtained, and the section weight set of each candidate path is constructed accordingly. and the total number of vehicles on the road segment ; The best path score is calculated by the best planning path formula for each candidate path's road segment weight set and the total number of vehicles on the road segment. , select the best path score The highest candidate path is taken as the best planning path; The optimal planning path formula is: ; in, is the number of road segments contained in the candidate path, Is the first candidate path The road weight of each road segment, Is the first candidate path The total number of vehicles on the road section, Is the first candidate path The rated maximum vehicle load capacity of each road section.

[0011] Preferably, the process of obtaining the driving fine-tuning action instruction is as follows: The participant's motion data is used as the input of the Transformer model and processed by the encoder and decoder to generate the participant's future trajectory coordinate sequence, namely the participant's predicted motion trajectory. The encoder contains a 6-layer self-attention mechanism, and the decoder contains a 6-layer cross-attention mechanism. Based on this, the participant's predicted motion trajectory and the vehicle speed, vehicle direction and vehicle acceleration in the vehicle status data are input into the deep Q network algorithm. Through reinforcement learning decision-making, driving fine-tuning action instructions including steering wheel steering angle and vehicle acceleration are obtained.

[0012] Preferably, the vehicle position is used as the starting point and the intersection coordinates are used as the end point. The straight-line distance between the two points is calculated using the Haversine formula as the vehicle intersection distance. The vehicle intersection distance is compared with a preset distance threshold. If the vehicle intersection distance is less than the preset distance threshold, an activation signal is sent to the road infrastructure collaboration module, triggering it to calculate the vehicle intersection trajectory based on the vehicle status data, traffic light timing data, and intersection geographic parameters. A collaboration instruction is also sent to the intersection comprehensive collaboration decision module, causing it to enter a standby state, ready to receive the vehicle intersection trajectory and perform dangerous spatiotemporal window detection. At the same time, through the instruction priority arbitration mechanism, the push of the driving fine-tuning action instruction generated by the participant interactive decision module to the vehicle execution control system is temporarily terminated, and the driving fine-tuning action instruction is automatically discarded. The preset distance threshold is dynamically changed according to the type of intersection. The preset distance threshold for urban intersections is 150m, the preset distance threshold for main road intersections is 300m, and the preset distance threshold for highway entrances and exits is 500m. The types of intersections include urban intersections, main road intersections and highway entrances and exits.

[0013] Preferably, the process of obtaining the trajectory of a vehicle passing through an intersection is as follows: The vehicle status data, traffic light timing data, and intersection geographical parameters including lane width, intersection boundary coordinates, and lane line positions are processed by the MPC algorithm. The MPC algorithm generates a position sequence with a fixed time interval through rolling optimization, i.e., the trajectory of the vehicle passing through the intersection.

[0014] Preferably, the dangerous space-time window is obtained by calculating the predicted motion trajectory of the participant and the trajectory of the vehicle through the intersection through the IoU algorithm. The dangerous space-time window is the time range and spatial position where the vehicle and the participant may collide, which is used to re-correct the trajectory of the vehicle through the intersection and provide a clear avoidance target.

[0015] Preferably, the process of correcting the trajectory of a vehicle passing through an intersection is as follows: The vehicle status data, traffic light timing data, and intersection geographic parameters including lane widths, intersection boundary coordinates, and lane line positions are processed through a constrained MPC algorithm to generate a corrected vehicle trajectory through the intersection. The constrained MPC algorithm is an MPC algorithm that uses the dangerous space-time window as a hard constraint. The dangerous space-time window is represented by a time interval and a spatial polygon. As a hard constraint condition that cannot be broken by the MPC algorithm, a constrained MPC algorithm is formed, wherein the time interval is the time range in which a collision may occur between the vehicle and the participant, and the spatial polygon is the spatial position covering the possible collision between the vehicle and the participant.

[0016] Due to the adoption of the above technical solution, the present invention has the following technical advancements compared to the prior art: 1. This invention breaks through the limitations of traditional static path planning and constructs a three-level planning mechanism consisting of a dynamic traffic network model, an improved A* algorithm, and multi-objective optimization. The GNN-LSTM fusion model is used to predict the average vehicle speed and congestion level of road sections in future time periods, enabling the traffic network model to reflect changes in road conditions in real time. The improved A* algorithm introduces a traffic flow balancing factor to dynamically balance single-route cost and global load during path search, preventing planning results from falling into local optimality. At the same time, multi-objective optimization achieves adaptive adjustment between traffic efficiency and traffic flow balance through quantitative analysis of the road section weight set and the total number of vehicles, effectively improving the overall traffic efficiency of the road network, reducing the load rate of congested sections, and solving the problems of delayed planning and aggravated congestion in traditional systems.

[0017] 2. This invention enhances the safety of participant interactions and the accuracy of intersection collaborative control. It uses the Transformer model to deeply mine the spatiotemporal correlation characteristics of participant action data, improves the accuracy of trajectory prediction for future time periods, and combines the deep Q-network algorithm to fuse the predicted trajectory with the vehicle state into a multi-dimensional state vector, generating precise fine-tuning instructions, shortening response delay and reducing collision risk. At the same time, it innovatively designs a dynamic trigger mechanism based on intersection type to accurately initiate collaborative decision-making, generate vehicle trajectories through intersections through the MPC algorithm, and combine the IoU algorithm to detect dangerous spatiotemporal windows. If there is a risk, the constrained MPC correction is initiated. This mechanism solves the defects of interactive decision-making lag and inappropriate intersection control timing in traditional systems, and achieves collaborative optimization of safety and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0019] Figure 1 It is a schematic diagram of the system function modules of the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0021] Examples, such as Figure 1 The invention relates to a vehicle-road collaborative path planning and decision-making system based on deep learning, which includes a data perception and acquisition module, a global path planning module, a participant interaction decision-making module, an intersection scenario triggering module, a road infrastructure collaborative module, an intersection comprehensive collaborative decision-making module and a vehicle execution control system.

[0022] The data sensing and acquisition module is used to obtain vehicle status data, vehicle travel destination, participant action data, global traffic flow data, traffic light timing data, road topology and intersection geographical parameters in real time; The global path planning module is used to build a traffic network model based on road topology and global traffic flow data, and to perform a path search on the traffic network model using an improved A* algorithm in combination with the vehicle's destination and location to obtain several candidate paths. The module then performs multi-objective optimization on the candidate paths to obtain the optimal planned path. The participant interaction decision module is used to obtain driving fine-tuning action instructions based on the participant action data and vehicle status data; The intersection scenario trigger module is used to calculate the vehicle intersection distance based on the vehicle position and intersection coordinates, compare the vehicle intersection distance with a preset distance threshold, determine whether to synchronously trigger the road infrastructure collaboration module and the intersection comprehensive collaborative decision module, and push the blocking driving fine-tuning action instruction to the vehicle execution control system; The road infrastructure collaboration module is used to obtain the trajectory of vehicles passing through the intersection based on vehicle status data, traffic light timing data and intersection geographical parameters; The intersection comprehensive collaborative decision-making module is used to calculate the dangerous time-space window of the vehicle's trajectory through the intersection, and if the dangerous time-space window exists, the vehicle's trajectory through the intersection is corrected; The vehicle execution control system is used to control the vehicle to execute the optimal planned path, as well as driving fine-tuning action instructions or the vehicle's trajectory through the intersection.

[0023] Furthermore, the working principle of the present invention is explained below by taking the entire process of an autonomous driving vehicle driving from the exit of a shopping mall parking lot to the entrance of an office building on an urban road network as an example: The data perception and acquisition module acquires various data in real time through multi-source fusion. Specifically, it uses lidar and millimeter-wave radar to collect participant motion data within a 360-degree range around the vehicle, such as pedestrian A's position time series (coordinates at t=0 (116.32°E, 39.91°N), at t=1s (116.3201°E, 39.91005°N), etc.), speed time series (0.8m / s→0.9m / s→…), and direction time series (due east→5°east→…). Vehicle status data is obtained through the on-board CAN bus, including vehicle position (116.319°E, 39.908°N), vehicle speed of 30km / h, vehicle direction due east, and vehicle acceleration of 0.2m / s. 2 Global traffic flow data and traffic light timing data are obtained from the roadside unit (RSU) through PC5 direct communication. The global traffic flow data includes the total number of vehicles on section L1 (from the shopping mall to intersection A), 20, a section length of 1000m, an average speed of 25km / h, and a congestion level of 0.2; and the total number of vehicles on section L2 (from intersection A to the office building), 15, a section length of 800m, an average speed of 30km / h, and a congestion level of 0.1. Traffic light timing data includes the current green light at intersection A, the remaining time of 18s, and the next three cycles of "green 30s → yellow 3s → The vehicle's destination (coordinates of the office building entrance (116.325°E, 39.912°N)), road topology (section L1 connects the starting point and intersection A, and section L2 connects intersection A and the end point) and intersection geographic parameters (intersection A is an urban intersection with a lane width of 3.5 m, intersection boundary coordinates (116.32°E±0.001°, 39.91°N±0.001°), the lane line is a through lane, and the intersection coordinates are the midpoint of the stop line (116.32°E, 39.91°N)) are obtained through the vehicle navigation system.

[0024] Construct a traffic network model and calculate the section weights of sections L1 and L2. According to the section weight formula, the section weight of section L1 is (1000m / 25km / h)×(1+0.5×0.2 2 )=40.8, the section weight of section L2 is: (800m / 30km / h)×(1+0.5×0.1 2)≈26.79, based on which the initial model of the traffic network is constructed. The initial model of the traffic network is analyzed and predicted by the GNN-LSTM fusion model. GNN extracts the spatial correlation between L1 and L2 (L1 is the upstream section of L2, and the traffic flow is conductive), and outputs the state vector of the fused spatial features. LSTM learns the historical 5-minute section traffic flow data and predicts the average speed of the section L1 and the congestion level of 0.25 in the next 5 minutes, and the average speed of the section L2 and the congestion level of 0.12. Based on this, the section weights of sections L1 and L2 are recalculated. The updated section weight of section L1 is (1000 / 22)×(1+0.5×0.25 2 )≈46.86, the updated weight of section L2 is (800 / 28)×(1+0.5×0.12 2 )≈28.77, and finally the traffic network model is obtained.

[0025] Improve the heuristic function of A* algorithm by introducing traffic flow balance factor , searching for paths from the traffic network model with the mall parking lot exit as the starting point and the office building entrance as the end point, and finding three candidate paths: Path 1: L1→L2, path cost value =46.86+28.77=75.63; Path 2: L1 → branch L3 → L2, path cost value 78.2, in the range [75.63, 75.63 × 1.3]; Path 3: L1 → branch L4 → L2, path cost 80.1, within the range [75.63, 75.63×1.3]; Split the three candidate paths into sections and calculate the best path score ( , ), the total weight of the road section of path 1 is 75.63, the total number of vehicles on the road section is [20,15], and the corresponding rated maximum vehicle load is [50,40]. The load balancing item The calculated value is 0.02, so the best path score of path 1 is 52.94. The best path scores of paths 2 and 3 are 51.2 and 49.8 respectively. Therefore, path 1 is selected as the best planned path.

[0026] The Transformer model inputs the position, speed, and direction time series of pedestrian A over the past 10 seconds. After processing through a 6-layer encoder and decoder, it predicts the predicted motion trajectory of pedestrian A for the next 5 seconds: (116.3202°E, 39.9101°N) at t=1s, (116.3203°E, 39.91015°N) at t=2s, etc., moving generally toward the southeast. The deep Q network inputs pedestrian A's predicted motion trajectory as well as the vehicle speed, direction, and acceleration from the vehicle status data. Through reinforcement learning, it makes decisions and outputs a driving fine-tuning action instruction: turn the steering wheel right. (avoid pedestrians), acceleration -0.3m / s 2 (Deceleration), because the distance between the vehicle's current position (116.3195°E, 39.9095°N) and the intersection coordinates (116.32°E, 39.91°N) is calculated by the Haversine formula to be 120m, and the intersection is an urban intersection, the preset distance threshold is 150m, so the road infrastructure coordination module and the intersection comprehensive coordination decision module are triggered simultaneously, and the above-mentioned "steering wheel 3°, acceleration -0.3m / s 2 " driving fine-tuning action instructions.

[0027] The vehicle status data, traffic light timing data and intersection geographical parameters are input into the MPC algorithm. The MPC algorithm generates a fixed time interval position sequence through rolling optimization: [(116.3195°E,39.9095°N),...,(116.32°E,39.91°N) (stop line),...,(116.3205°E,39.91°N)], which is the trajectory of the vehicle passing through the intersection. It means that in the next 15 seconds, it will drive at a constant speed from the current straight lane, reach the stop line (within the green light period) at 10 seconds, and completely pass through the intersection and enter the L2 section at 15 seconds. The U algorithm calculates the overlap between pedestrian A's predicted trajectory and the vehicle's intersection trajectory, yielding an Intersection of Unions (IoU) of 0.35. This indicates a dangerous spatiotemporal window (12-14 seconds in time and at a spatial location of (116.32°E±0.0005°, 39.91°N±0.0005°). The constrained MPC algorithm (using the dangerous spatiotemporal window as a hard constraint) corrects the vehicle's intersection trajectory by delaying the stop line by 2 seconds (reaching the stop line at 12 seconds), adjusting the vehicle speed to 25 km / h, and avoiding the danger zone to ensure it passes through the overlapping space, i.e., the dangerous spatiotemporal window, after 14 seconds.

[0028] The vehicle is controlled to prioritize executing the corrected intersection passing trajectory. After passing the intersection, it continues to travel along the optimal planned path L1→L2 until it reaches the destination.

[0029] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A vehicle-road cooperative path planning decision system based on deep learning, characterized by: include: Data perception and acquisition module, used to obtain real-time vehicle status data, vehicle destination, participant action data, global traffic flow data, traffic light timing data, road topology and intersection geographical parameters; The global path planning module is used to build a traffic network model based on road topology and global traffic flow data. It then uses an improved A* algorithm to search for paths in the traffic network model based on the vehicle's destination and location, obtaining several candidate paths. The module then performs multi-objective optimization on these candidate paths to obtain the optimal planned path. Participant interaction decision module, used to obtain driving fine-tuning action instructions based on participant action data and vehicle status data; The intersection scenario trigger module is used to calculate the vehicle intersection distance based on the vehicle position and intersection coordinates, compare the vehicle intersection distance with the preset distance threshold, determine whether to synchronously trigger the road infrastructure collaboration module and the intersection comprehensive collaborative decision module, and push the blocking driving fine-tuning action instruction to the vehicle execution control system; The road infrastructure collaboration module is used to obtain the trajectory of vehicles passing through intersections based on vehicle status data, traffic light timing data, and intersection geographical parameters; The intersection comprehensive collaborative decision-making module is used to calculate the dangerous time-space window of the vehicle's trajectory through the intersection. If the dangerous time-space window exists, the vehicle's trajectory through the intersection is corrected; The vehicle execution control system is used to control the vehicle to execute the optimal planned path, as well as fine-tune driving action instructions or the vehicle's trajectory through intersections.

2. The deep learning-based vehicle-road cooperative path planning decision system according to claim 1, characterized in that: The data perception and acquisition module obtains vehicle status data and participant action data in real time through on-board sensors, and obtains global traffic flow data and traffic light timing data through V2X and the vehicle's driving destination, road topology and intersection geographical parameters through the on-board navigation system interface. The vehicle status data includes vehicle position, vehicle speed, vehicle direction and vehicle acceleration, the participant action data includes participant position timing sequence, participant speed timing sequence and participant direction timing sequence, the global traffic flow data includes the total number of vehicles on each road section, the section length, the average speed and congestion level, and the intersection geographical parameters include the type of intersection, the width of each lane at the intersection, the intersection boundary coordinates, the lane line position and the intersection coordinates.

3. The deep learning-based vehicle-road cooperative path planning decision system according to claim 2, characterized in that: The method for constructing a traffic network model based on road topology and global traffic flow data: S1. Calculate the segment weight of each segment using the segment weight formula based on the segment length, average vehicle speed, and congestion level of each segment in the global traffic flow data. Based on this, use the segment weight of each segment as the edge weight of the corresponding edge in the road topology to obtain an initial traffic network model. S2. Analyze and predict the initial traffic network model using the GNN-LSTM fusion model to obtain the predicted average speed and congestion level for each road segment in the initial traffic network model, that is, the predicted average speed and congestion level for each edge. S3. The predicted average vehicle speed, predicted congestion level, and length of each road section are processed again through S1 to obtain a traffic network model.

4. The deep learning-based vehicle-road cooperative path planning decision system according to claim 3, characterized in that: The process of performing path search on the traffic network model by improving the A* algorithm: The improved A* algorithm takes the vehicle location as the starting point and the vehicle destination as the end point, searches the path of the traffic network model, and obtains the path cost value. Several candidate paths within the range, among which, is the optimal candidate path cost value, is the preset cost value threshold; The improved A* algorithm introduces a traffic flow balance factor into the heuristic function A* algorithm.

5. The deep learning-based vehicle-road cooperative path planning decision system according to claim 4, characterized in that: The method for performing multi-objective optimization on a plurality of candidate paths: Performing segment-level splitting on each of the candidate routes to obtain each route segment, the total number of vehicles on the corresponding route segment, and the route segment weight, thereby forming a route segment weight set and a route segment total number set for each candidate route; The best path score is calculated for each candidate path's road section weight set and the total number of vehicles on the road section using the best planning path formula, and the candidate path with the highest best path score is selected as the best planning path.

6. The deep learning-based vehicle-road cooperative path planning decision system according to claim 5, characterized in that: The process of obtaining the driving fine-tuning action instruction: The participant's motion data is input into the Transformer model for processing to obtain the participant's predicted motion trajectory. The participant's predicted motion trajectory and the vehicle speed, vehicle direction and vehicle acceleration in the vehicle status data are then processed through a deep Q-network algorithm to obtain driving fine-tuning action instructions including steering wheel angle and vehicle acceleration.

7. The deep learning-based vehicle-road cooperative path planning decision system according to claim 6, characterized in that: The vehicle intersection distance is compared with a preset distance threshold. If the vehicle intersection distance is less than the preset distance threshold, the road infrastructure coordination module and the intersection comprehensive coordination decision module are synchronously triggered, and the blocking driving fine-tuning action instruction is pushed to the vehicle execution control system; The preset distance threshold is dynamically changed according to the type of intersection. The preset distance threshold for urban intersections is 150m, the preset distance threshold for main road intersections is 300m, and the preset distance threshold for highway entrances and exits is 500m. The types of intersections include urban intersections, main road intersections and highway entrances and exits.

8. The deep learning-based vehicle-road cooperative path planning decision system according to claim 7, characterized in that: The process of obtaining the trajectory of the vehicle passing through the intersection: The vehicle status data, traffic light timing data, and intersection geographical parameters including lane width, intersection boundary coordinates, and lane line positions are processed through the MPC algorithm to obtain the trajectory of the vehicle passing through the intersection.

9. The deep learning-based vehicle-road cooperative path planning decision system according to claim 8, characterized in that: The dangerous spatiotemporal window is obtained by calculating the predicted motion trajectory of the participant and the trajectory of the vehicle passing through the intersection through the IoU algorithm.

10. The deep learning-based vehicle-road cooperative path planning decision system according to claim 9, characterized in that: The process of correcting the trajectory of a vehicle passing through an intersection: The vehicle status data, traffic light timing data, and intersection geographic parameters including lane widths, intersection boundary coordinates, and lane line positions are processed through a constrained MPC algorithm to generate a corrected vehicle trajectory through the intersection. The constrained MPC algorithm is an MPC algorithm that uses the dangerous time-space window as a hard constraint.

Images