A deep learning-based vehicle-road cooperation path planning decision system

Abstract

The application discloses a kind of vehicle-road cooperation path planning decision system based on deep learning, belong to intelligent transportation technical field.The system is by collecting global traffic flow data, vehicle position and vehicle driving destination, with GNN-LSTM fusion model carries out each road section traffic flow prediction, provides basis for global path planning;Utilize the trajectory of participant motion predicted by Transformer model, realize the real-time interactive decision of vehicle and participant by DQN algorithm;Based on MPC algorithm, the vehicle state data, traffic light timing data and intersection geographic parameters are fused to generate the vehicle trajectory through the intersection, and the IoU algorithm is used to detect whether the dangerous space-time window exists, to determine whether to use the MPC algorithm with constraints to correct the vehicle trajectory through the intersection, to realize the intersection collaborative decision.The application improves the dynamic adaptability of vehicle path planning, enhances the vehicle-road cooperation and the interaction safety with surrounding participants, and effectively improves the traffic efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent transportation, and particularly relates to a vehicle-road cooperative path planning and decision system based on deep learning. BACKGROUND

[0002] With the rapid development of intelligent transportation technology, vehicle path planning and vehicle-road cooperation have become the core research direction to improve traffic efficiency and safety. In the prior art, the vehicle path planning system is mostly based on static road information or local traffic data for decision-making, which has the problems of insufficient dynamic adaptability, weak vehicle-road cooperation ability and low safety in interaction with surrounding participants, and is difficult to meet the needs of complex traffic scenarios.

[0003] In terms of path planning, the traditional system relies on static road network model and fixed weight algorithm, which cannot respond to global traffic flow dynamic changes in real time, is easy to fall into local optimum, and leads to the decline of overall traffic efficiency; in terms of vehicle-road cooperation and participant interaction, the trajectory prediction accuracy of surrounding participants is insufficient, and the historical regularity and spatial correlation are not deeply mined, leading to delayed or misjudged driving fine-tuning decisions, and the cooperation between vehicles and road infrastructure is mostly one-way information receiving, lacking dynamic trajectory cooperation based on real-time vehicle state and signal timing, and being easy to appear "green light empty" or "red light rush".

[0004] In summary, the prior art has obvious deficiencies in dynamic traffic adaptation, deep vehicle-road cooperation, participant safety interaction and intersection precise decision-making, and there is an urgent need for a vehicle-road cooperative path planning system that integrates deep learning technology and has global perception and cooperative decision-making capability to improve traffic efficiency and driving safety. SUMMARY

[0005] The application aims to provide a vehicle-road cooperative path planning and decision system based on deep learning to solve the problems raised in the background.

[0006] To solve the above technical problems, the application provides the following technical solutions:

[0007] A vehicle-road cooperative path planning and decision system based on deep learning, comprising a data perception and collection module, a global path planning module, a participant interaction decision module, an intersection scenario triggering module, a road infrastructure cooperation module, an intersection comprehensive cooperation decision module and a vehicle execution control system.

[0008] The data-aware acquisition module is configured to acquire vehicle state data, a vehicle travel destination, participant action data, global traffic flow data, traffic light timing data, road topology, and intersection geographical parameters in real time.

[0009] The global path planning module is configured to construct a traffic network model based on the road topology and the global traffic flow data, and to search for a path in the traffic network model by using an improved A* algorithm in combination with the vehicle travel destination and the vehicle position, to obtain a plurality of candidate paths, and to perform multi-objective optimization processing on the plurality of candidate paths to obtain a best planning path.

[0010] The participant interaction decision module is configured to obtain a travel fine-tuning action instruction based on the participant action data and the vehicle state data.

[0011] The intersection scenario triggering module is configured to calculate a vehicle intersection distance based on the vehicle position and intersection coordinates, to compare the vehicle intersection distance with a preset distance threshold, and to determine whether to synchronously trigger the road infrastructure cooperation module and the intersection comprehensive cooperation decision module and to block the travel fine-tuning action instruction from being pushed to the vehicle execution control system.

[0012] The road infrastructure cooperation module is configured to obtain a vehicle trajectory through an intersection based on the vehicle state data, the traffic light timing data, and the intersection geographical parameters.

[0013] The intersection comprehensive cooperation decision module is configured to calculate a dangerous space-time window of the vehicle trajectory through the intersection, and to correct the vehicle trajectory through the intersection if the dangerous space-time window exists.

[0014] The vehicle execution control system is configured to control the vehicle to execute the best planning path, the travel fine-tuning action instruction, or the vehicle trajectory through the intersection.

[0015] Preferably, the data-aware acquisition module serves as a basic data input terminal, comprehensively perceives traffic scene information through multi-source fusion, and obtains vehicle state data and participant action data in real time through vehicle-mounted sensors, including laser radar, millimeter wave radar, camera, wheel speed sensor, acceleration sensor and steering angle sensor. The laser radar and millimeter wave radar are responsible for collecting the three-dimensional position coordinates and relative speed of participants including pedestrians and other vehicles within a 360-degree range around the vehicle in real time. The camera extracts the motion direction and posture features of the participants through image recognition technology, and after synchronization by time stamp, forms continuous position time sequence, speed time sequence and direction time sequence of the participants, constituting the participant action data. At the same time, the vehicle state data is obtained in real time through the vehicle-mounted CAN bus, wherein the vehicle position is output by the GPS positioning module combined with the RTK technology, the vehicle speed and vehicle acceleration are measured jointly by the wheel speed sensor and the acceleration sensor, and the vehicle direction is obtained by fusion calculation of the electronic compass and the steering angle sensor.

[0016] V2X interacts with road side unit (RSU) through PC5 direct communication to obtain global traffic flow data and traffic light timing data, wherein the global traffic flow data covers the entire traffic network and includes the total number of vehicles on each road segment, the length of the road segment, the average vehicle speed and the congestion degree. The traffic light timing data includes the signal light state, the remaining time of the signal light state and the signal light timing plan for the next three periods.

[0017] The vehicle navigation system interface is connected to the map service platform through OTA to obtain the latitude and longitude coordinates of the vehicle driving destination, the road topology and the intersection geographical parameters including the intersection type, the width of each lane at the intersection, the boundary coordinates of the intersection, the position of the lane line and the intersection coordinates, wherein the intersection coordinates are the latitude and longitude coordinates of the midpoint of the stop line at the intersection, serving as the reference point for triggering intersection scenario decision.

[0018] Preferably, the specific method for constructing a traffic network model based on road topology and global traffic flow data is as follows:

[0019] S1. Calculate the road segment weight of each road segment in the global traffic flow data by the road segment weight formula, wherein the road segment length , the average vehicle speed and the congestion degree of each road segment are used as the road segment weight of each road segment, the road topology is used to form an initial topology network with intersections as nodes and road segments between intersections as edges, and the road segment weight of each road segment is used as the edge weight of the corresponding edge in the road topology to obtain an initial model of the traffic network.

[0020] The road segment weight formula is as follows:

[0021] ;

[0022] wherein, is a regulation parameter;

[0023] S2. The initial traffic network model is analyzed and predicted by the GNN-LSTM fusion model, the initial traffic network model is regarded as a graph structure, the nodes are intersections, the edges are road segments, and the edge features include the average speed and congestion degree of each road segment , and the spatial correlation between road segments is extracted by the graph neural network GNN, and the road segment state vector with fused spatial features is output, the LSTM inputs the road segment state vector output by the GNN into the LSTM layer in time sequence, learns the time evolution law of the road segment traffic flow, and finally outputs the predicted average speed and predicted congestion degree of each road segment;

[0024] S3. The predicted average speed and predicted congestion degree of each road segment output by S2 and the corresponding road segment length are re-substituted into the road segment weight formula in S1, and the updated road segment weight of each road segment is calculated, and the updated road segment weight is re-assigned to the edge weight of the road topology corresponding edge, and the traffic network model is obtained.

[0025] Preferably, the process of searching for a path for the traffic network model by improving the A* algorithm is as follows:

[0026] The heuristic function of the improved A* algorithm is based on the heuristic function of the A* algorithm, and a traffic flow balance factor is introduced.

[0027] The heuristic function of the improved A* algorithm is as follows:

[0028] ;

[0029] wherein, is a comprehensive cost evaluation value, is the actual cost from the starting point to the current node which is an intermediate point or intersection of a road segment, is the estimated cost from the current node to the end point, is a regulation coefficient, and the value is ;

[0030] The improved A* algorithm takes the vehicle position as the starting point and the vehicle driving destination as the end point, and performs a depth-first search on the traffic network model to search for a path. When the first candidate path, i.e., the combination of continuous road segments from the starting point to the end point, is searched, the path cost value of the first candidate path is calculated and defined as the optimal candidate path cost value , wherein the path cost value of the first candidate path is the sum of the road segment weights of all road segments in the path.

[0031] ​After the first candidate path search is completed, other candidate paths are searched, and only paths with path cost values in the range of are retained as candidate paths, wherein is a preset cost value threshold and has a value of If the number of candidate paths searched is less than 3, the threshold is automatically relaxed to 0.5.

[0032] Preferably, the method for multi-objective optimization processing of a plurality of candidate paths comprises:

[0033] Each candidate path in the plurality of candidate paths is split at the road segment level to obtain each road segment included in each candidate path, and the corresponding total number of road segment vehicles and road segment weight, and the total number of road segment vehicles and the road segment weight set of each candidate path are constructed according to the road segment weight set ;

[0034] The road segment weight set and the total number of road segment vehicles of each candidate path are calculated by the best planning path formula to obtain the best path score , and the candidate path with the highest best path score is selected as the best planning path.

[0035] The best planning path formula is:

[0036] ;

[0037] wherein is the number of road segments included in the candidate path, is the road segment weight of the th road segment in the candidate path, is the total number of road segment vehicles of the th road segment in the candidate path, is the rated vehicle maximum carrying capacity of the th road segment in the candidate path.

[0038] Preferably, the process of obtaining the driving fine-tuning action instruction comprises:

[0039] The participant action data is input into a Transformer model, and the participant's future trajectory coordinate sequence, i.e., the participant's predicted motion trajectory, is generated by processing the input through an encoder and a decoder, wherein the encoder comprises 6 layers of self-attention mechanism, and the decoder comprises 6 layers of cross-attention mechanism. The participant's predicted motion trajectory and the vehicle speed, vehicle direction and vehicle acceleration in the vehicle state data are input into a deep Q network algorithm, and the driving fine-tuning action instruction including the steering angle of the steering wheel and the vehicle acceleration is obtained through reinforcement learning decision.

[0040] ​​​Preferably, the straight-line distance between the vehicle position and the intersection coordinate is calculated by the Haversine formula as the vehicle intersection distance, and the vehicle intersection distance is compared with the 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 to trigger the calculation of the vehicle passing intersection trajectory based on the vehicle state data, the traffic light timing data, and the intersection geographic parameters, and a collaboration instruction is sent to the intersection comprehensive collaboration decision module to make it enter the standby state, prepare to receive the vehicle passing intersection trajectory and perform the dangerous spatiotemporal window detection, and temporarily terminate the push of the driving fine-tuning action instruction generated by the participant interactive decision module to the vehicle execution control system through the instruction priority arbitration mechanism, and the driving fine-tuning action instruction is automatically discarded.

[0041] The preset distance threshold is dynamically changed according to the intersection type. The preset distance threshold of the urban intersection is 150 m, the preset distance threshold of the main road intersection is 300 m, and the preset distance threshold of the highway exit is 500 m. The intersection type includes an urban intersection, a main road intersection, and a highway exit.

[0042] Preferably, the process of obtaining the vehicle passing intersection trajectory includes the following steps:

[0043] The vehicle state data, the traffic light timing data, and the intersection lane width, the intersection boundary coordinate, and the lane line position in the intersection geographic parameters are processed by the MPC algorithm. The MPC algorithm generates a position sequence with a fixed time interval, i.e., the vehicle passing intersection trajectory, through rolling optimization.

[0044] Preferably, the dangerous spatiotemporal window is obtained by the IoU algorithm based on the participant predicted motion trajectory and the vehicle passing intersection trajectory. The dangerous spatiotemporal window is the time range and the spatial position in which the vehicle and the participant may collide, and is used to provide a clear avoidance target for the re-correction of the vehicle passing intersection trajectory.

[0045] Preferably, the process of correcting the vehicle passing intersection trajectory includes the following steps:

[0046] The vehicle state data, the traffic light timing data, and the intersection geographic parameters including the intersection lane width, the intersection boundary coordinate, and the lane line position are processed by the MPC algorithm with constraints to generate the corrected vehicle passing intersection trajectory.

[0047] The MPC algorithm with constraints is an MPC algorithm with the dangerous spatiotemporal window as a hard constraint. The dangerous spatiotemporal window is represented in the form of a time interval and a spatial polygon, and is used as a hard constraint condition that cannot be broken by the MPC algorithm, forming the MPC algorithm with constraints. The time interval is the time range in which the vehicle and the participant may collide, and the spatial polygon is the spatial position in which the vehicle and the participant may collide.

[0048] Due to the adoption of the above technical scheme, the present application has the following technical progress compared with the prior art:

[0049] 1. The present application breaks through the limitations of traditional static path planning, constructs a three-level planning mechanism of traffic network dynamic model, improved A* algorithm and multi-objective optimization, predicts the average speed and congestion degree of road segments in the future period through the GNN-LSTM fusion model, so that the traffic network model can reflect the road condition changes in real time, and through the improved A* algorithm, the traffic flow balancing factor is introduced to dynamically balance the single road cost and global load in path search, avoiding the planning result from falling into local optimum, while the multi-objective optimization realizes adaptive adjustment between traffic efficiency and traffic flow balance through quantitative analysis of road segment weight set and vehicle total number set, effectively improving the overall traffic efficiency of road network, reducing the load rate of congested road segments, and solving the problems of lagging planning and aggravating congestion of traditional system.

[0050] 2. The present application enhances the safety of participant interaction and the accuracy of intersection cooperative control, uses the Transformer model to deeply mine the spatio-temporal correlation features of participant action data, improves the accuracy of trajectory prediction in the future period, and combines the deep Q network algorithm to fuse the predicted trajectory and vehicle state into a multi-dimensional state vector to generate accurate fine-tuning instructions, shorten the response delay and reduce the collision risk; at the same time, a dynamic trigger mechanism based on intersection type is innovatively designed to accurately start cooperative decision-making, generate vehicle trajectory through the MPC algorithm, and detect dangerous spatio-temporal window through the IoU algorithm, and if there is a risk, start the MPC repair with constraints, which solves the defects of lagging interaction decision-making and improper intersection control timing of traditional systems, and realizes the cooperative optimization of safety and efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0052] Figure 1 is a schematic diagram of the system function modules of the present application. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0054] An embodiment, as described above, is a deep learning-based vehicle-road cooperative path planning and decision system, which comprises a data perception and collection module, a global path planning module, a participant interaction decision module, a road intersection scenario triggering module, a road infrastructure cooperation module, a road intersection comprehensive cooperation decision module, and a vehicle execution control system. Figure 1

[0055] The data perception and collection module is configured to acquire vehicle state data, a vehicle driving destination, participant action data, global traffic flow data, traffic light timing data, road topology, and road intersection geographic parameters in real time.

[0056] The global path planning module is configured to construct a traffic network model based on road topology and global traffic flow data, and to search for a path for the traffic network model by using an improved A* algorithm in combination with a vehicle driving destination and a vehicle position, to obtain a plurality of candidate paths, and to perform multi-objective optimization processing on the plurality of candidate paths to obtain a best planning path.

[0057] The participant interaction decision module is configured to obtain a driving fine-tuning action instruction based on participant action data and vehicle state data.

[0058] The road intersection scenario triggering module is configured to calculate a vehicle-road intersection distance based on a vehicle position and a road intersection coordinate, to compare the vehicle-road intersection distance with a preset distance threshold, to determine whether to synchronously trigger the road infrastructure cooperation module and the road intersection comprehensive cooperation decision module, and to block the driving fine-tuning action instruction from being pushed to the vehicle execution control system.

[0059] The road infrastructure cooperation module is configured to obtain a vehicle trajectory through a road intersection based on vehicle state data, traffic light timing data, and road intersection geographic parameters.

[0060] The road intersection comprehensive cooperation decision module is configured to calculate a dangerous space-time window of the vehicle trajectory through the road intersection, and to correct the vehicle trajectory through the road intersection if the dangerous space-time window exists.

[0061] The vehicle execution control system is configured to control the vehicle to execute the best planning path, the driving fine-tuning action instruction, or the vehicle trajectory through the road intersection.

[0062] Further, the working principle of the present application is described below by taking a whole process of an autonomous vehicle driving from an exit of a parking lot of a shopping mall to an entrance of an office building in a city road network as an embodiment.

[0063] ​The data-aware collection module acquires various types of data in real time through multi-source fusion. Specifically, the motion data of participants within a 360-degree range around the vehicle is collected by laser radar and millimeter wave radar, such as the position time sequence of pedestrian A (coordinates (116.32 °E, 39.91 °N) at t=0, (116.3201 °E, 39.91005 °N) at t=1 s, …), the speed time sequence (0.8 m / s→0.9 m / s→…), and the direction time sequence (due east→east 5° south→…); vehicle state data is acquired through the vehicle CAN bus, including vehicle position (116.319 °E, 39.908 °N), vehicle speed 30 km / h, vehicle direction due east, and vehicle acceleration 0.2 m / s 2 ; global traffic flow data and traffic light timing data are acquired from the road side unit (RSU) through PC5 direct communication, wherein the global traffic flow data includes the total number of vehicles on road segment L1 (from the shopping mall to intersection A) 20, the length of the road segment 1000 m, the average speed 25 km / h, and the congestion degree 0.2, and the total number of vehicles on road segment L2 (from intersection A to the office building) 15, the length of the road segment 800 m, the average speed 30 km / h, and the congestion degree 0.1; the traffic light timing data includes the current green light at intersection A, the remaining time 18 s, and the future 3 cycles as a "green 30 s→yellow 3 s→red 40 s" cycle; the vehicle driving destination (the office building entrance coordinates (116.325 °E, 39.912 °N)), the road topology (road segment L1 connects the starting point and intersection A, and road segment L2 connects intersection A and the terminal point), and the intersection geographical parameters (intersection A is a city intersection, the intersection lane width is 3.5 m, the intersection boundary coordinates are (116.32 °E±0.001 °, 39.91 °N±0.001 °), the lane line is a straight lane, and the intersection coordinates are the midpoint of the stop line (116.32 °E, 39.91 °N)) are acquired through the vehicle navigation system.

[0064] A traffic network model is constructed to calculate the road segment weights of road segment L1 and road segment L2. According to the road segment weight formula, the road segment weight of road segment L1 is (1000 m / 25 km / h)×(1+0.5×0.2 2 )=40.8, and the road segment weight of road segment L2 is (800 m / 30 km / h)×(1+0.5×0.1 2)≈26.79, according to which the initial model of the traffic network is constructed, and the initial model of the traffic network is analyzed and predicted by the GNN-LSTM fusion model, the GNN extracts the spatial correlation of L1 and L2 (L1 is the upstream road section of L2, and the traffic flow has a conduction property), and outputs the state vector of the fused spatial features, the LSTM learns the historical 5-minute road section traffic flow data, and predicts the average speed of 22 km / h and the congestion degree of 0.25 of the road section L1 in the next 5 minutes, and the average speed of 28 km / h and the congestion degree of 0.12 of the road section L2, according to which the road section weights of the road section L1 and the road section L2 are recalculated, the updated road section weight of the road section L1 is (1000 / 22) x (1+0.5x0.25 2 )≈46.86, and the updated road section weight of the road section L2 is (800 / 28) x (1+0.5x0.12 2 )≈28.77, and finally the traffic network model is obtained.

[0065] The traffic flow equilibrium factor is introduced into the heuristic function of the improved A* algorithm The path is searched from the parking lot exit of the shopping mall as the starting point and the office building entrance as the ending point in the traffic network model, and three candidate paths are searched:

[0066] Path 1: L1→L2, path cost value =46.86+28.77=75.63;

[0067] Path 2: L1→branch L3→L2, path cost value 78.2, in the range of [75.63, 75.63x1.3];

[0068] Path 3: L1→branch L4→L2, path cost value 80.1, in the range of [75.63, 75.63x1.3];

[0069] The three candidate paths are split at the road section level, and the best path score is calculated (Pbest) , ), the total road section weight of path 1 is 75.63, the total number of road section vehicles is [20, 15], and the corresponding maximum rated vehicle carrying capacity is [50, 40], and the load balancing term is calculated as 0.02, according to which the best path score of path 1 is 52.94, and the best path scores of paths 2 and 3 are 51.2 and 49.8 respectively, so path 1 is selected as the best planning path.

[0070] The Transformer model inputs the position time sequence, speed time sequence and direction time sequence of pedestrian A in the past 10 seconds, processes through 6 layers of encoder and decoder, and predicts the future 5 seconds of pedestrian A predicted motion trajectory: t=1s (116.3202°E, 39.9101°N), t=2s (116.3203°E, 39.91015°N), …, moving in the southeast direction as a whole; the deep Q network inputs the pedestrian A predicted motion trajectory and the vehicle speed, vehicle direction and vehicle acceleration in the vehicle state data, makes a decision through reinforcement learning, and outputs the driving fine-tuning action instruction: steering wheel right turn (evasive pedestrian), acceleration -0.3 m / s 2 (deceleration), because the current position of the vehicle (116.3195°E, 39.9095°N), the distance from the intersection coordinate (116.32°E, 39.91°N) is calculated by the Haversine formula as 120m, and the preset distance threshold is 150m, so the road infrastructure cooperation module and the intersection comprehensive cooperation decision module are triggered synchronously, and the above driving fine-tuning action instruction of “steering wheel 3°, acceleration -0.3 m / s 2 ” is blocked.

[0071] The vehicle state data, traffic light time sequence data and intersection geographic parameters are input into the MPC algorithm, the MPC algorithm generates a fixed time interval position sequence [(116.3195°E, 39.9095°N), …, (116.32°E, 39.91°N) (stop line), …, (116.3205°E, 39.91°N)] through rolling optimization, that is, the vehicle passing through the intersection trajectory, which represents that the vehicle will uniformly drive on the current straight lane in the next 15s, reach the stop line at 10s (in the green light period), and enter the L2 section after 15s. According to this, the IoU algorithm is used to calculate the overlap degree of the pedestrian A predicted motion trajectory and the vehicle passing through the intersection trajectory, and the IoU=0.35 is obtained. There is a dangerous spatiotemporal window: time range 12-14s, spatial position (116.32°E±0.0005°, 39.91°N±0.0005°)), and a constrained MPC algorithm (regarding the dangerous spatiotemporal window as a hard constraint) is used to modify the vehicle passing through the intersection trajectory: delaying 2s to pass through the stop line (arriving at the stop line at 12s), adjusting the vehicle speed to 25km / h, and the trajectory avoids the dangerous area to ensure that it passes through the overlapping space after 14s, that is, the dangerous spatiotemporal window.

[0072] The control vehicle preferentially executes the modified intersection passing trajectory, and after passing through the intersection, continues to drive along the optimal planning path L1→L2 until reaching the terminal point.

[0073] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A deep learning-based vehicle-road cooperation path planning decision system, characterized in that, The application relates to a vehicle intelligent driving system and method. The application comprises: a data-aware acquisition module for acquiring vehicle state data, vehicle driving destination, participant action data, global traffic flow data, traffic light timing data, road topology and intersection geographical parameters in real time; a global path planning module for constructing a traffic network model based on the road topology and the global traffic flow data, and searching for a path of the traffic network model through an improved A* algorithm in combination with the vehicle driving destination and the vehicle position, so as to obtain a plurality of candidate paths, and performing multi-objective optimization processing on the plurality of candidate paths to obtain a best planning path; a participant interaction decision module for obtaining driving fine-tuning action instructions according to the participant action data and the vehicle state data; an intersection scenario triggering module for calculating a vehicle intersection distance according to the vehicle position and intersection coordinates, comparing the vehicle intersection distance with a preset distance threshold, and judging whether to synchronously trigger a road infrastructure cooperation module and an intersection comprehensive cooperation decision module and block the driving fine-tuning action instructions from being pushed to a vehicle execution control system; a road infrastructure cooperation module for obtaining a vehicle intersection trajectory according to the vehicle state data, the traffic light timing data and the intersection geographical parameters; an intersection comprehensive cooperation decision module for calculating a dangerous space-time window of the vehicle intersection trajectory, and modifying the vehicle intersection trajectory if the dangerous space-time window exists; a vehicle execution control system for controlling the vehicle to execute the best planning path, the driving fine-tuning action instructions or the vehicle intersection trajectory; the vehicle intersection distance is compared with the preset distance threshold, and if the vehicle intersection distance is smaller than the preset distance threshold, the road infrastructure cooperation module and the intersection comprehensive cooperation decision module are synchronously triggered, and the driving fine-tuning action instructions are blocked from being pushed to the vehicle execution control system; the preset distance threshold is dynamically changed according to intersection types, the preset distance threshold of an urban intersection is 150 m, the preset distance threshold of a main road intersection is 300 m, and the preset distance threshold of a highway exit is 500 m, wherein the intersection types include an urban intersection, a main road intersection and a highway exit; the process of obtaining the vehicle intersection trajectory comprises the following steps: vehicle state data, traffic light timing data and intersection geographical parameters including intersection lane width, intersection boundary coordinates and lane line position are processed through an MPC algorithm to obtain the vehicle intersection trajectory; the dangerous space-time window is calculated through an IoU algorithm from a participant predicted motion trajectory and the vehicle intersection trajectory; The improved A* algorithm takes the vehicle position as the starting point and the vehicle driving destination as the end point to perform path searching on the traffic network model to obtain a path cost value a plurality of candidate paths within the range, wherein the optimal candidate path cost value, a preset cost value threshold; The improved A* algorithm introduces a traffic flow equilibrium factor in the heuristic function of the A* algorithm; the process of searching for a path of the traffic network model through the improved A* algorithm comprises the following steps: the process of modifying the vehicle intersection trajectory comprises the following steps: vehicle state data, traffic light timing data and intersection geographical parameters including intersection lane width, intersection boundary coordinates and lane line position are processed through a constrained MPC algorithm to generate a modified vehicle intersection trajectory; the constrained MPC algorithm is an MPC algorithm taking the dangerous space-time window as a hard constraint.

2. The deep learning-based vehicle-road cooperation path planning decision system according to claim 1, wherein, The data-aware collection module acquires vehicle state data and participant action data in real time through vehicle-mounted sensors, and acquires global traffic flow data and traffic light timing data through V2X, and acquires vehicle driving destination, road topology and intersection geographic parameters through a vehicle-mounted navigation system interface, wherein the vehicle state data includes vehicle position, vehicle speed, vehicle direction and vehicle acceleration, the participant action data includes participant position time sequence, participant speed time sequence and participant direction time sequence, the global traffic flow data includes total number of vehicles on each road segment, road segment length, average vehicle speed and congestion degree, and the intersection geographic parameters include intersection type, intersection lane width, intersection boundary coordinates, lane line position and intersection coordinates. 3.The deep learning based vehicle-road cooperative path planning and 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 comprises the following steps: S1. The road segment length, average vehicle speed and congestion degree of each road segment in the global traffic flow data are used to calculate the road segment weight of each road segment through a road segment weight formula, and the road segment weight of each road segment is used as the edge weight of the corresponding edge in the road topology to obtain an initial traffic network model; S2. The initial traffic network model is analyzed and predicted through a GNN-LSTM fusion model to obtain the predicted average vehicle speed and predicted congestion degree of each road segment in the initial traffic network model, i.e. the predicted average vehicle speed and predicted congestion degree of each edge; S3. The predicted average vehicle speed and predicted congestion degree of each road segment and the road segment length are reprocessed through S1 to obtain a traffic network model.

4. The deep learning-based vehicle-road cooperation path planning decision system according to claim 3, characterized in that, The method for multi-objective optimization processing of a plurality of candidate paths comprises the following steps: Each candidate path is split at the road segment level to obtain each road segment contained in each candidate path and the corresponding total number of vehicles and road segment weight, and a road segment weight set and a total number of vehicles set of each candidate path are formed accordingly; The road segment weight set and the total number of vehicles set of each candidate path are used to calculate the best path score through a best planning path formula, and the candidate path with the highest best path score is selected as the best planning path.

5. The deep learning-based vehicle-road cooperation path planning decision system according to claim 4, characterized in that, The process for obtaining driving fine-tuning action instructions comprises the following steps: The participant action data is input into a Transformer model for processing to obtain a predicted participant motion trajectory, and the predicted participant motion trajectory and the vehicle speed, vehicle direction and vehicle acceleration in the vehicle state data are processed through a deep Q network algorithm to obtain driving fine-tuning action instructions including steering angle and vehicle acceleration.

Images