Vehicle collaborative decision-making system and method based on real-time traffic situation prediction

Abstract

The invention discloses a vehicle collaborative decision-making system and method based on real-time traffic situation prediction, and relates to the technical field of intelligent traffic. According to the invention, multi-source traffic data are fused through a vehicle-road-cloud cooperative sensing system, and a real-time dynamic traffic scene map is constructed; based on the improved LSTM neural network, traffic situation prediction is carried out by introducing space-time map convolution and a double attention mechanism; the prediction result is converted into risk field and opportunity field data; a hierarchical optimization strategy of centralized planning and distributed execution is adopted, a global optimization instruction is generated based on a genetic algorithm, each vehicle carries out local adjustment according to risk field and opportunity field data inquired in real time, a final control instruction is generated, and multi-vehicle cooperative control with both safety and efficiency is achieved. According to the method, the problems of incomplete perception, inaccurate prediction, delayed control response and the like in a complex traffic environment are effectively solved, and the overall efficiency and safety of regional traffic are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent transportation, and in particular, relates to a vehicle cooperative decision-making system and method based on real-time traffic situation prediction. BACKGROUND

[0002] With the rapid development of smart cities and intelligent networked vehicles, the transportation system is evolving from traditional isolated control to "vehicle-road-cloud" integrated and cooperative control. Globally, governments and industries are vigorously promoting cooperative sensing and decision-making technologies based on V2X, high-precision maps, and multi-source perception, aiming to solve long-standing problems such as road congestion, frequent accidents, and low efficiency.

[0003] In typical complex scenarios such as urban intersections, highway ramp merging areas, and construction road sections, the behaviors of traffic participants strongly influence each other, with a large number of blind spots, mixed conflicts, and uncertainties. Traditional single-vehicle intelligence, due to limited perception and isolated decision-making, is difficult to cope with sudden intrusions, pedestrian disturbances, and multi-vehicle gaming, leading to local congestion chain reactions and even collision accidents. How to achieve global optimization and safety control through multi-vehicle cooperation has become a core challenge to improve regional traffic efficiency and safety.

[0004] In the prior art, prediction models based on fixed rules or traditional machine learning cannot effectively express the complex spatio-temporal topological relationships between traffic participants. For example, LSTM networks cannot model the interactive influence between vehicles, resulting in insufficient prediction accuracy. At the same time, most cooperative control methods rely on centralized scheduling, with high response delay and poor flexibility, and do not quantify risks and traffic opportunities as computable fields, making control strategies lack foresight and adaptability. In addition, the cooperative mechanism between global optimization and distributed execution is imperfect, making it difficult to balance efficiency and safety in dynamic scenarios. SUMMARY

[0005] (I) Technical problems solved To solve the problems in the related art, the present application provides a vehicle cooperative decision-making system and method based on real-time traffic situation prediction to overcome the above technical problems existing in the prior art.

[0006] (II) Technical solutions To solve the above technical problems, the present application is realized by the following technical solutions: S1, collect vehicle data, roadside data, and cloud data and fuse them to obtain multi-source traffic fusion data; based on the multi-source traffic fusion data, construct a real-time dynamic traffic scene graph; S2, based on the time sequence and spatial characteristics of the real-time dynamic traffic scene graph, introduce a spatio-temporal graph convolution network and an attention mechanism in the LSTM neural network to obtain an initial LSTM traffic situation prediction network; The initial LSTM traffic situation prediction network is trained using the historical dynamic traffic scene graph with interaction event labels to obtain a final LSTM traffic situation prediction network; and the real-time dynamic traffic scene graph is input into the final LSTM traffic situation prediction network to obtain traffic situation prediction data; S3, converting the traffic situation prediction data into risk field data and opportunity field data; S4, based on the risk field data and the opportunity field data, a hierarchical optimization strategy of centralized planning combined with distributed execution is used to obtain real-time vehicle control instructions; The real-time vehicle control instructions are distributed to the vehicles in a collaborative instruction distribution manner for collaborative control; The present application realizes the establishment of a digital twin base by fusing car-road-cloud multi-source data to construct a real-time dynamic traffic scene graph, which comprehensively and accurately perceives the traffic environment; the improved LSTM neural network is used for situation prediction, the spatiotemporal graph convolution and double attention mechanism are introduced to accurately predict the vehicle trajectory and interaction behavior, and the deep understanding of the complex traffic scene is realized; the prediction data is converted into risk field and opportunity field to quantitatively evaluate the future spatiotemporal risk and opportunity, and accurate navigation is provided for vehicle control; the hierarchical strategy of centralized planning and distributed execution is adopted to optimize the global traffic efficiency and realize the flexible and safe control of individual vehicles; the industry pain points such as perception blind area, inaccurate prediction and control lag are effectively solved, and the safety, efficiency and collaboration level of the traffic system are significantly improved.

[0007] Preferably, the S1 comprises the following steps: S11, collecting the global position, speed, acceleration, heading angle, turn signal state and planning path of all collaborative vehicles in the target area to obtain vehicle data; S12, collecting road geometry information, traffic signal state and countdown, pedestrian / non-motor vehicle position and blind area supplementary information to obtain roadside data; S13, obtaining macro traffic flow data including regional average speed, congestion index and event information from a traffic cloud platform to obtain cloud data; S14, aligning all data to a unified timestamp and coordinate system by Kalman filtering based on the vehicle data, roadside data and cloud data to obtain multi-source traffic fusion data; S15, constructing a static traffic graph layer; constructing a real-time dynamic traffic scene graph based on the multi-source traffic fusion data and the static traffic graph layer; Preferably, the S15 comprises the following steps: S151, constructing a static traffic graph layer; the static traffic graph layer comprises road level information and semantic level information; S152, taking the static traffic graph layer as a bottom plate, based on the multi-source traffic fusion data, creating a dynamic node for each participant, creating a relationship edge between the nodes according to the spatial position and the map spectrum semantics, and obtaining a real-time dynamic traffic scene graph; The application forms a real-time dynamic traffic scene graph by constructing a static traffic graph layer as a bottom plate and creating dynamic nodes and relationship edges on the basis.

[0008] Preferably, the S2 comprises the following steps: S21, based on the time sequence and spatial characteristics of the dynamic traffic scene graph, an initial LSTM traffic situation prediction network is obtained by introducing a spatio-temporal graph convolution network and an attention mechanism into the LSTM neural network and improving the LSTM neural network; S22, collecting historical dynamic traffic scene graphs; key interaction events in the historical dynamic traffic scene graphs are labeled, and each event is labeled with a start time and a vehicle involved, to obtain a historical dynamic traffic scene graph with interaction events; S23, the initial LSTM traffic situation prediction network is trained using the historical dynamic traffic scene graph with interaction events, and the optimal learning rate of the initial LSTM traffic situation prediction network is found by an optimization algorithm in the training process to obtain an optimal solution; the optimal solution is used as the learning rate of the initial LSTM traffic situation prediction network to obtain a final LSTM traffic situation prediction network; S24, the real-time dynamic traffic scene graph is input into the final LSTM traffic situation prediction network to obtain traffic situation prediction data; The application improves the LSTM neural network, introduces a spatio-temporal graph convolution and a double attention mechanism, and trains and optimizes using labeled historical data to construct a high-precision traffic situation prediction model; accurate prediction of vehicle trajectories and interaction behaviors is realized, and the situation awareness capability in complex traffic scenes is effectively improved, providing a reliable basis for collaborative decision-making.

[0009] Preferably, the S21 comprises the following steps: S211, replacing the fully connected layers in the input gate, the forget gate and the output gate in the LSTM neural network with a graph convolution gate unit; S212, introducing a dynamic graph attention mechanism in the graph convolution gate unit for calculating the aggregation weight of the neighbor nodes; S213, introducing a time sequence self-attention mechanism after the stacked graph convolution gate unit for processing the hidden state sequence output by the graph convolution gate unit; The improved traffic situation prediction network is constructed by replacing the gating fully connected layer of the LSTM with a graph convolution unit, and introducing dynamic graph attention and timing self-attention mechanisms respectively, so that the model can capture the spatial interaction relationship and long time sequence dependence between traffic participants at the same time, realize more accurate prediction of vehicle trajectory and interaction behavior in complex traffic scenes, and significantly improve the accuracy and reliability of situation prediction.

[0010] Preferably, the training process in S23 finds the optimal learning rate of the initial LSTM traffic situation prediction network through an optimization algorithm, and the optimal solution includes the following steps: S231, set the prediction accuracy and resource occupancy rate of the initial LSTM traffic situation prediction network; set the maximum optimization iteration number, and set the survival space size of the artificial fish swarm; S232, based on the survival space size of the artificial fish swarm, an initial artificial fish swarm set is constructed, and the size of the initial artificial fish swarm set is set according to the learning rate of the initial LSTM traffic situation prediction network; S233, according to the prediction accuracy and resource occupancy rate, define the fitness function of the yield rate brought by the number of artificial fish swarms; S234, perform iterative operation on the initial artificial fish swarm set, the higher the fitness value, the more optimal the number of artificial fish swarms; in each iteration process, the fitness value of the number of artificial fish swarms is calculated according to the fitness function, the number of artificial fish swarms is updated from high to low according to the fitness value, and the current best artificial fish swarm number and the global best artificial fish swarm number in the artificial fish swarm are obtained in each iteration process; S235, repeat S234, when the maximum optimization iteration number is reached, stop iteration, and take the global best fish swarm number as the optimal solution; The learning rate of the LSTM network is optimized by the artificial fish swarm algorithm, a fitness function with prediction accuracy and resource occupancy rate as the target is constructed, and the optimal solution is obtained through iterative search; the best balance between accuracy and efficiency in the model training process is realized, the shortcomings of traditional manual parameter tuning are effectively avoided, and the convergence speed and generalization ability of the traffic situation prediction model are significantly improved, providing a stable and reliable model basis for real-time prediction.

[0011] Preferably, S3 includes the following steps: S31, divide the target area into a regular two-dimensional grid; assign a unique spatial coordinate to each grid cell in the two-dimensional grid, and establish a mapping relationship with the static elements in the real-time dynamic traffic scene graph to obtain a two-dimensional grid with a mapping relationship; S32, dividing a prediction time period in the traffic situation prediction data into equal-interval time slices to obtain a time slice set; assigning a time stamp t to each time slice in the time slice set to obtain a complete time sequence from a current time to a prediction end point; mapping the two-dimensional grid along the complete time sequence from the current time to the prediction end point to obtain a three-dimensional space-time grid structure; S33, for each pair of interacting vehicles, calculating a conflict probability of each point in the space-time grid based on a prediction trajectory distribution in the traffic situation prediction data to obtain conflict probability data; obtaining a covariance matrix of each prediction trajectory from the final LSTM traffic situation prediction network to obtain a representation of prediction uncertainty; generating multiple possible trajectories based on the representation of prediction uncertainty through Monte Carlo sampling, calculating a probability variance of each space-time grid unit in the space-time grid being occupied by a vehicle to obtain an uncertainty contribution; weighting and fusing the conflict risk and the uncertainty risk, and smoothing and diffusing in the space-time dimension through a Gaussian kernel function to obtain risk field data; S34, identifying a continuous region not occupied by a vehicle in the space-time grid to obtain a passing gap set; collecting a spatial size, a time window and reachability of each gap in the passing gap set to obtain space-time gap data; constructing a gap quality scoring model based on a safety margin, stability and practicability; processing the space-time gap data through the gap quality scoring model to calculate a quality score of each gap in the passing gap set to obtain a gap quality score set; assigning the gap quality score in the gap quality score set to a corresponding space-time grid unit to form an initial opportunity distribution; filtering the initial opportunity distribution based on vehicle dynamics constraints to remove gaps that cannot be utilized due to physical limitations to obtain a processed opportunity distribution; applying a space-time diffusion operation to make a high-quality gap in the processed opportunity distribution have a positive impact on its adjacent space-time region to obtain opportunity field data; The application converts the traffic situation prediction data into quantifiable risk field and opportunity field by constructing a three-dimensional space-time grid structure; calculates a conflict probability and an uncertainty risk based on a prediction trajectory, fuses to generate a risk field; simultaneously identifies a passing gap and evaluates its quality to form an opportunity field; realizes accurate quantification and visual expression of future traffic risk and passing opportunity, effectively solves the problem that risk and opportunity are difficult to quantify in traditional methods, and provides space-time navigation information with safety and high efficiency for vehicle cooperative decision-making.

[0012] Preferably, the S4 comprises the following steps: S41, based on traffic situation prediction data, a global optimization problem is constructed, the optimization problem aims to minimize the total travel time, and a genetic algorithm is used to find the optimal global instruction of the global optimization problem; S42, the global optimal global instruction is issued to each vehicle, each vehicle executes the target speed in the global optimal instruction as the basic execution speed, establishes the basic road right relationship according to the traffic priority, and obtains the initial real-time control instruction; The risk threshold and opportunity threshold are set, the vehicle real-time queries the risk field value and opportunity field value on the current position and predicted trajectory based on the risk field and opportunity field data, and the real-time vehicle control instruction is obtained based on the risk field value and opportunity field value and in combination with the initial real-time control instruction; S43, the real-time vehicle control instruction is issued to the vehicle in a cooperative instruction distribution manner for cooperative control; The present application realizes the cooperative control of centralized planning and distributed execution by using the genetic algorithm to generate the global optimization instruction and combining the distributed dynamic fine-tuning of the vehicle based on the risk field and opportunity field data, ensures the optimal global traffic efficiency, gives the individual vehicle flexibility to deal with real-time risks, effectively solves the problem of response lag of traditional centralized scheduling, and significantly improves the safety and travel efficiency of vehicle cooperation in complex traffic environment.

[0013] Preferably, the step of finding the optimal solution of the global optimization problem in S41 by using the genetic algorithm comprises the following steps: S411, decision variables of the global optimization problem are set, the decision variables comprise a travel sequence and a recommended speed, and a fitness function of the global optimization problem is constructed; S412, an initial chromosome population is constructed, each chromosome of the chromosome population represents a travel sequence and a recommended speed of a group of vehicles, and a maximum optimization iteration number is set; S413, iteration is started, the fitness of each chromosome in the chromosome population is calculated according to a second fitness function formula in each iteration process, a chromosome fitness set is obtained, the chromosomes in the chromosome population are selected according to the fitness set, a selected chromosome population is obtained, and the chromosomes in the selected chromosome population are subjected to crossover and mutation operations to obtain an operated chromosome population; S414, S413 is repeated, and the optimal global instruction is obtained when the maximum optimization iteration number is reached; The present application globally optimizes the vehicle travel sequence and recommended speed by using the genetic algorithm, constructs a fitness function that comprehensively considers the travel delay and safety risk, iteratively solves the optimal instruction by selecting, crossing and mutating, realizes the overall optimal planning of regional traffic efficiency, effectively avoids the local optimization trap, and significantly improves the road travel efficiency while ensuring safety.

[0014] A vehicle cooperative decision system based on real-time traffic situation prediction is used to realize the vehicle cooperative decision method based on real-time traffic situation prediction, comprising a data collection and fusion module, a traffic situation prediction module, a risk and opportunity field generation module and a vehicle control instruction generation and cooperative control module. The data collection and fusion module is used to collect traffic data from multiple sources in real time and perform integrated processing; by obtaining vehicle data, road side data and cloud data; all data are aligned to a unified timestamp and coordinate system by using Kalman filtering, forming multi-source traffic fusion data, and based on this, a real-time dynamic traffic scene graph is constructed, which includes a static layer and a dynamic layer, and is updated at a fixed period to provide a basis for subsequent analysis; The traffic situation prediction module is based on the real-time dynamic traffic scene graph, and uses an improved LSTM neural network to predict the traffic situation; by introducing a dynamic graph attention mechanism and a time sequence self-attention mechanism in the graph convolution LSTM unit, the understanding ability of the model for spatio-temporal interaction features is enhanced, so as to construct an initial LSTM traffic situation prediction network; the network is trained using a historical dynamic traffic scene graph with interaction event annotation, and the learning rate is optimized by an artificial fish swarm algorithm to obtain a final prediction network; the real-time dynamic traffic scene graph is input into the network, and traffic situation prediction data including vehicle future trajectory and interaction event probability are output; The risk and opportunity field generation module converts the traffic situation prediction data into risk field and opportunity field data to quantify the dangers and opportunities in the traffic environment; the target area is divided into a three-dimensional spatio-temporal grid structure, then the conflict probability and uncertainty risk are calculated, and the risk field data is obtained by weighted fusion; the passing gap in the spatio-temporal grid is identified, the safety margin, stability and practicability are evaluated, the gap quality score is generated, and the opportunity field data is formed by diffusion operation; these field data represent the risk level and opportunity availability in the spatio-temporal dimension, providing a basis for vehicle control; The vehicle control instruction generation and cooperative control module generates real-time vehicle control instructions based on the risk field and opportunity field data through a hierarchical optimization strategy; at the global level, an optimization problem is constructed to minimize the total travel time, and a genetic algorithm is used to assign passing priority and recommended speed to obtain optimal global instructions; at the vehicle level, the basic execution speed is adjusted according to the real-time queried risk field and opportunity field values to generate the final control instructions; the instructions are distributed to vehicles through a cooperative instruction distribution method to realize cooperative control of multiple vehicles and improve overall traffic efficiency and safety.

[0015] (Three) beneficial effects The present application has the following beneficial effects: The application significantly improves the comprehensiveness and accuracy of traffic environment perception through multi-source data fusion and dynamic graph construction; the application utilizes a vehicle-road-cloud collaborative perception system, integrates vehicle state, roadside perception and macro traffic flow and other multi-source heterogeneous data, and performs time and space alignment through Kalman filtering to construct a real-time scene graph containing static traffic elements and dynamic participants; the graph accurately represents complex traffic environment and interaction in a unified digital model, providing a reliable and complete data basis for subsequent prediction and decision-making, effectively overcoming the single-vehicle intelligent perception blind area problem.

[0016] The application realizes high-precision and forward-looking prediction of traffic situation based on an improved deep learning prediction model; the application innovatively introduces a spatio-temporal graph convolution network and a double attention mechanism into LSTM, enabling the model to not only capture the time sequence changes of traffic flow, but also understand the spatial topological relationship and interaction between vehicles; this design makes traffic situation prediction, especially the prediction of key interactive behaviors such as lane changing and conflict, more accurate and consistent with reality, providing key forward-looking information for risk assessment and control decisions.

[0017] The application innovatively quantizes the prediction results into a risk field and an opportunity field, providing a decision basis for vehicle control that combines safety and efficiency; the application generates an intuitive risk field by calculating conflict probability and uncertainty risk in a three-dimensional spatio-temporal grid; the application constructs an opportunity field by identifying and evaluating passing gaps; the application converts abstract prediction data into calculable and queryable spatio-temporal distribution information, enabling vehicles to simultaneously perceive future dangerous areas and passing opportunities, thereby achieving efficiency optimization under the premise of safety.

[0018] The application adopts a hierarchical optimization strategy of centralized planning and distributed execution, achieving an effective balance between global efficiency and individual flexibility; the application performs global optimization in the cloud through a genetic algorithm to develop a macro passing sequence and recommended speed, ensuring regional overall traffic efficiency; the application allows each vehicle to fine-tune global instructions based on locally real-time queried risk field and opportunity field data; this collaborative mechanism not only leverages the systematic advantages of centralized planning, but also retains the rapid response capabilities of distributed execution, ultimately achieving safe, smooth and efficient collaborative control in complex dynamic scenarios.

[0019] Of course, any product implementing the application does not necessarily need to achieve all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed for the description of the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and all other drawings obtained by those of ordinary skill in the art without creative effort based on these drawings also belong to the protection scope of the present application.

[0021] Figure 1 A flowchart of a vehicle cooperative decision-making method based on real-time traffic situation prediction of the present application; Figure 2 A module schematic diagram of a vehicle cooperative decision-making system based on real-time traffic situation prediction of the present application; Figure 3 A flowchart of constructing a real-time dynamic traffic scene graph in a vehicle cooperative decision-making method based on real-time traffic situation prediction of the present application; Figure 4 A flowchart of obtaining a final LSTM traffic situation prediction network in a vehicle cooperative decision-making method based on real-time traffic situation prediction of the present application; Figure 5 A structure schematic diagram of an LSTM traffic situation prediction network in a vehicle cooperative decision-making method based on real-time traffic situation prediction of the present application; Figure 6 A flowchart of obtaining opportunity field data in a vehicle cooperative decision-making method based on real-time traffic situation prediction of the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings of the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, but not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0023] Please refer to Figure 1 The embodiments of the present application disclose a vehicle cooperative decision-making method based on real-time traffic situation prediction, comprising the following steps: S1, collecting vehicle data, roadside data and cloud data and fusing to obtain multi-source traffic fusion data; based on the multi-source traffic fusion data, a real-time dynamic traffic scene graph is constructed; The S1 comprises the following steps: S11, through the vehicle CAN bus and the V2X communication module, the global position (GPS / Beidou), speed, acceleration, heading angle, steering light state and planning path of all cooperative vehicles in the target area are collected in real time to obtain vehicle data; S12, collect road geometry information, traffic signal state and countdown, pedestrian / non-motor vehicle position and blind area supplementary information through roadside units (RSUs) and perception facilities (cameras, millimeter wave radars), and obtain roadside data; S13, obtain macroscopic traffic flow data including regional average speed, congestion index and event information (accidents, construction) from a traffic cloud platform, and obtain cloud data; S14, align all data to a unified timestamp and coordinate system through Kalman filtering of vehicle data, roadside data and cloud data, and obtain multi-source traffic fusion data; S15, construct a static traffic graph layer; construct a real-time dynamic traffic scene graph based on multi-source traffic fusion data and the static traffic graph layer; Please refer to Figure 3 , the S15 includes the following steps: S151, construct a static traffic graph layer; the static traffic graph layer includes road level information and semantic level information; the road level information includes lane lines, lane center lines, stop lines and flow guide areas; the semantic level information includes lane types, traffic signs (speed limits, yielding) and intersection connection relationships (which lane can lead to which lane); S152, take the static traffic graph layer as a base plate, create a dynamic node for each participant (vehicle, pedestrian) based on multi-source traffic fusion data, create relationship edges between nodes according to spatial positions and map spectrum semantics, and obtain a real-time dynamic traffic scene graph; the dynamic node includes ID, type, position, speed and acceleration; the relationship edge includes a spatial relationship edge and a semantic relationship edge; the spatial relationship edge is, for example, vehicle A is on lane L1 and pedestrian P is on zebra crossing Z; the interaction relationship edge is, for example, vehicle B is the front vehicle of vehicle C (based on the relative position in the lane) and vehicle D and vehicle E have potential path conflicts (based on the predicted trajectory and intersection geometry calculation); The real-time dynamic traffic scene graph runs in a fixed cycle (for example, 100 milliseconds); in each cycle, the system repeatedly performs S11 to S15 to update the state of all nodes and edges in the dynamic scene graph with the latest data; nodes that are out of date or leave the area will be removed, and new nodes will be created; S2, based on the timing and spatial characteristics of the real-time dynamic traffic scene graph, introduce a spatiotemporal graph convolution network and an attention mechanism in an LSTM neural network, and obtain an initial LSTM traffic situation prediction network; An initial LSTM traffic situation prediction network is trained using a historical dynamic traffic scene graph with interaction event annotations, and a final LSTM traffic situation prediction network is obtained; a real-time dynamic traffic scene graph is input into the final LSTM traffic situation prediction network, and traffic situation prediction data is obtained; Referring to Figure 4 、 Figure 5 , the S2 comprises the following steps: S21, based on the time sequence and space characteristics of the dynamic traffic scene graph, an initial LSTM traffic situation prediction network is obtained by introducing a spatio-temporal graph convolution network and an attention mechanism into the LSTM neural network and improving the LSTM neural network; The S21 comprises the following steps: S211, replacing the fully connected layer in the input gate, the forget gate and the output gate in the LSTM neural network with a graph convolution gate unit; The input of the traditional LSTM is a flat feature vector, which can meet the time sequence characteristics in the dynamic traffic data in the application, but cannot effectively express the topological relationship between vehicles in the traffic scene; for example, it cannot understand which vehicle is close to which vehicle, which vehicle is behind which vehicle, and it regards each vehicle as an isolated individual; a model that can understand "spatial context" is needed; for example, the braking behavior of the front vehicle has a huge impact on the following vehicle, but has little impact on the vehicles in the adjacent lane or far away; therefore, a graph convolution layer that understands spatial features is needed; Specifically, the matrix multiplication (full connection) in the standard LSTM is replaced by a graph convolution operation, and creating a graph convolution LSTM unit includes graph structure construction and application of graph convolution in LSTM gating; The graph structure construction constructs a graph for each time of traffic scene, and the graph includes nodes and edges; each vehicle is a node, and the node feature hi is the state vector of the vehicle, such as [x coordinate, y coordinate, speed, acceleration, heading angle]; if two vehicles are within a certain spatial distance (or according to the lane relationship of the high-precision map), an edge is established between them, defining the interaction range of the vehicles; The application of the graph convolution in the LSTM gating uses a graph convolution kernel to replace the traditional matrix multiplication; when each vehicle node state is updated, the information of the neighbor nodes (i.e. adjacent vehicles) is aggregated through the graph convolution operation, so that the model learns to understand the interaction rules based on the spatial topology, such as "the front vehicle decelerates, and the vehicle also needs to decelerate", solving the problem that the traditional LSTM cannot effectively encode the mutual influence between traffic participants; S212, introducing a dynamic graph attention mechanism into the graph convolution gate unit for calculating the aggregation weight of the neighbor nodes; The basic graph convolution usually assigns the same or predefined weights (e.g., based on distance) to all neighbor nodes of a node; but in reality, the influence is uneven; for example, in a lane-changing scenario, your influence weight on the car behind the target lane is much greater than that on the car in front of the same lane; therefore, the model needs to dynamically and adaptively determine which neighbor node has the greatest influence on the future trajectory of the self at a specific moment; Specifically, for each neighbor vehicle of the center vehicle, an attention coefficient is calculated, which is determined by the hidden states of the center vehicle and the neighbor vehicle; the coefficients are normalized using a softmax function to obtain attention weights representing the importance of interaction; when aggregating information, the features of the neighbor vehicles are weighted and summed according to the weights, rather than being treated equally; so that the model can adaptively focus on key influencers, such as paying more attention to the car behind the target lane than the car in front of the same lane when changing lanes, greatly improving the accuracy of prediction in complex interaction scenarios; S213, a temporal self-attention mechanism is introduced after the stacked graph convolution gated unit, which is used to process the hidden state sequence output by the graph convolution gated unit; although LSTM designs a gate to alleviate the long-term dependency problem, when the input sequence is long, it will gradually forget the information far in the past; in a traffic scenario, a "cut-in" or "hard braking" event a few seconds ago may have a lasting impact on the future; at this time, the model needs a mechanism that can directly review the entire historical sequence and autonomously determine which information at which time is crucial, so a temporal self-attention mechanism needs to be introduced; Specifically, a Transformer encoder is connected after the stacked graph convolution LSTM layer; self-attention operations are applied to the entire historical hidden state sequence output by the LSTM; for each historical time in the sequence, a correlation weight relative to all other times is calculated, so that distant but critical events can be directly captured; such as identifying a chance cut-in behavior a few seconds ago that has a key impact on the current time trajectory; this mechanism gives the model the ability to look at the entire history, so that its prediction is no longer dependent on recent states, but is based on a deep understanding of the entire causal chain of events, significantly improving the rationality of long-term prediction; S22, collect historical dynamic traffic scenario graphs; the collection process is the same as S1; key interaction events in the historical dynamic traffic scenario graphs are labeled, and each event is labeled with a start time and involved vehicles to obtain a historical dynamic traffic scenario graph with interaction events; the key interaction events are, for example, lane changing, following, intersection conflict, yielding; S23, training the initial LSTM traffic situation prediction network using the historical dynamic traffic scenario graph with interaction events, the training process finding the optimal learning rate of the initial LSTM traffic situation prediction network through an optimization algorithm to obtain an optimal solution; taking the optimal solution as the learning rate of the initial LSTM traffic situation prediction network to obtain a final LSTM traffic situation prediction network; The training process in S23 finds the optimal learning rate of the initial LSTM traffic situation prediction network through an optimization algorithm to obtain an optimal solution, including the following steps: S231, setting a prediction accuracy of the initial LSTM traffic situation prediction network p 1 and a resource occupancy rate is p 2; setting a maximum optimization iteration number and setting a survival space size of an artificial fish swarm; S232, constructing an initial artificial fish swarm set based on the survival space size of the artificial fish swarm, setting a size of the initial artificial fish swarm set as b , the initial artificial fish swarm set being represented as , wherein, u i , wherein, the initial artificial fish swarm set is represented as i , wherein, the initial artificial fish swarm set is represented as The survival space size of each artificial fish in the initial artificial fish swarm set is fixed, so the number of artificial fish will affect the size of the final adult fish, and too many or too few artificial fish will reduce the yield rate of the overall artificial fish swarm; S233, defining a fitness function of the yield rate brought by the number of artificial fish according to the prediction accuracy p 1 and the resource occupancy rate p 2, the fitness function formula being as follows, ; q , wherein, the fitness function is represented as l , wherein, the fitness function bias is used to adjust the output of the fitness function; S234, performing an iteration operation on the initial artificial fish swarm set, the higher the fitness value, the better the number of artificial fish; in each iteration process, the fitness value of the number of artificial fish is calculated according to the fitness function, the number of artificial fish is updated from high to low according to the fitness value, and the current best number of artificial fish in the artificial fish swarm and the global best number of artificial fish are obtained in each iteration process; S235, repeating S234, when the maximum optimization iteration number is reached, stopping the iteration and taking the global best number of fish as the optimal solution; S24, inputting the real-time dynamic traffic scenario graph into the final LSTM traffic situation prediction network to obtain traffic situation prediction data; S3, transforming the traffic situation prediction data into risk field data and opportunity field data; The S3 includes the following steps: S31, dividing a target area (such as an intersection or a highway section) into a regular two-dimensional grid, and the grid resolution is usually 0.5-1 meter, ensuring that the vehicle contour and gap can be accurately captured; Assigning a unique spatial coordinate to each grid cell in the two-dimensional grid, and establishing a mapping relationship with the static elements such as lanes and curbs in the real-time dynamic traffic scene graph, to obtain a two-dimensional grid with a mapping relationship; S32, dividing the prediction period (such as 8 seconds) in the traffic situation prediction data into equally spaced time segments, and the time resolution is usually 0.1-0.5 seconds, matching the vehicle control frequency to obtain a time segment set; assigning a time stamp t to each time segment in the time segment set to obtain a complete time sequence from the current time to the prediction endpoint; Obtaining a three-dimensional space-time grid structure along the complete time sequence from the current time to the prediction endpoint based on the two-dimensional grid with a mapping relationship; each grid cell (x, y, t) in the three-dimensional space-time grid structure represents a state container at a specific location at a specific time; S33, for each pair of interacting vehicles, calculating the conflict probability of each point in the space-time grid based on the predicted trajectory distribution in the traffic situation prediction data to obtain conflict probability data; the calculation formula is as follows, Pc(x, y, t) = Σi* ∑j*P (vi at (x,y,t)) x P(vj at (x,y,t)) x overlap area(i,j); where Pc(x,y,t) represents the collision probability of the interacting vehicles at (x, y,t), and the collision probability field; ∑i Σj represents double summation for each pair of interacting vehicles (i, j), where i≠j (excluding the same vehicle); (x, y, t) represents the coordinate point in the space-time grid, where (x, y) represents the two-dimensional spatial position and t represents the time point; i and j represent the index number of the traffic participants, i and j represent different vehicles; P(viat (x, y, t)) represents the probability value of vehicle i being at position (x, y) at time t, which is derived from the trajectory distribution output of the prediction model; P(vj at (x, y, t)) represents the probability value of vehicle j being at position (x, y) at time t; overlap area(i,j) represents the contour overlap area of vehicle i and vehicle j at position (x, y), considering the actual physical size of the vehicle; Obtaining the covariance matrix of each predicted trajectory from the final LSTM traffic situation prediction network to obtain the prediction uncertainty; based on the prediction uncertainty, multiple possible trajectories are generated by Monte Carlo sampling, and the probability variance of each space-time grid cell in the space-time grid being occupied by a vehicle is calculated to obtain the uncertainty contribution; the calculation formula is as follows, Ru(x, y, t) = Σk*Var[P(vk at (x, y, t))]; wherein, Ru(x, y, t) represents the uncertainty contribution of all vehicles at (x, y, t), i.e. the uncertainty risk field; k represents the index number of the traffic participant; ∑k represents the summation of all vehicles k, accumulating the uncertainty contribution of each vehicle; P(vk at (x, y, t)) represents the probability value of vehicle k being at position (x, y) at time t; Var[·] represents the variance operator, indicating the uncertainty degree of the probability prediction; Var[P(vk at (x, y, t))] represents the variance of the probability of vehicle k being at (x, y, t), reflecting the prediction confidence; The collision risk and the uncertainty risk are weighted and fused, and are smoothed and diffused in the space-time dimension through a Gaussian kernel function to obtain risk field data; the weighted fusion formula is as follows, Rf(x, y, t) = α × Pc(x, y, t) + β × Ru(x, y, t); wherein, Rf(x, y, t) represents the fused risk field data, the final risk evaluation value at the space-time grid (x, y, t); α represents the weight coefficient of the collision risk, taking a value in the range [0, 1], controlling the importance of the collision probability in the total risk; β represents the weight coefficient of the uncertainty risk, taking a value in the range [0, 1], controlling the importance of the prediction uncertainty in the total risk; Please refer to Figure 6 S34, identifying a continuous region in the space-time grid that is not occupied by a vehicle to obtain a passing gap set; collecting the spatial size, time window and reachability of each gap in the passing gap set to obtain space-time gap data; the spatial size represents the available size of the gap in the length and width directions; the time window represents the duration of the gap from appearing to disappearing; the reachability is based on the current speed and position, whether the vehicle can reach the gap within the time window; A gap quality scoring model is constructed based on safety margin, stability and practicability; the safety margin represents the safety distance of the gap from surrounding vehicles, the stability represents the continuous stability of the gap within the prediction period, and the practicability represents whether the gap meets the current driving intention of the vehicle (such as lane changing or merging); the gap quality scoring model is used to process the space-time gap data to calculate the quality score of each gap in the passing gap set to obtain a gap quality score set; Specifically, the gap quality assessment model scores the safety margin, stability and practicability of the passing gap; for the safety margin, the real-time dynamic distance between the gap and all surrounding traffic participants is calculated to ensure that it meets the minimum spatial buffer required for safe passing; for the stability, the spatio-temporal evolution law of the gap within the prediction period is analyzed to evaluate its duration and form retention ability, avoiding the selection of fleeting gaps; for the practicability, based on the current driving target of the vehicle (such as lane changing, merging or accelerating) and its trajectory planning, the matching degree of the gap and the expected path is judged; based on the cooperative evaluation of the three types of indicators, the system gives each gap a standardized quality score in the range of 0 to 1; The gap quality score in the gap quality score set is assigned to the corresponding spatio-temporal grid cell to form an initial opportunity distribution; based on the vehicle dynamics constraints, the initial opportunity distribution is filtered for reachability, removing gaps that cannot be utilized due to physical limitations, to obtain a processed opportunity distribution; The spatio-temporal diffusion operation is applied to make the high-quality gaps in the processed opportunity distribution have a positive impact on their adjacent spatio-temporal regions, obtaining opportunity field data; S4, based on the risk field data and the opportunity field data, a hierarchical optimization strategy of centralized planning combined with distributed execution is used to obtain real-time vehicle control instructions; The real-time vehicle control instructions are distributed to the vehicles through the way of cooperative instruction distribution for cooperative control; The S4 includes the following steps: S41, based on the traffic situation prediction data, a global optimization problem is constructed, the optimization problem takes the minimization of the total travel time of the entire region as the target, assigns the passing priority, the recommended speed interval and the macroscopic path to the vehicle; the genetic algorithm is used to find the optimal global instruction of the global optimization problem; The S41 includes the following steps: S411, set the decision variables of the global optimization problem; the decision variables include the passing sequence and the recommended speed; construct the fitness function of the global optimization problem; the formula of the second fitness function is as follows, F=-(j1×Total Delay + j2×Total Risk); Wherein, F represents the second fitness value; Total Delay represents the total travel delay, the total travel delay =∑(actual travel time of vehicle-ideal travel time of vehicle); the Total Risk represents the total safety risk, the total safety risk is obtained by summing up the safety risk of each vehicle; j1, j2 represent the weight coefficients of the total travel delay and the total safety risk, respectively; S412, construct an initial chromosome population; each chromosome of the chromosome population represents a passing sequence and a recommended speed of a group of vehicles; set a maximum number of optimization iterations; S413, start iteration, calculate the fitness of each chromosome in the chromosome population according to the second fitness function formula in each iteration process, obtain a chromosome fitness set; select chromosomes in the chromosome population according to the fitness set, obtain a selected chromosome population; perform crossover and mutation operations on the chromosomes in the selected chromosome population, obtain an operated chromosome population; S414, repeat S413, when the maximum number of optimization iterations is reached, obtain the optimal global instruction; S42, issue the global optimal global instruction to each vehicle; each vehicle takes the target speed in the global optimal instruction as the basic execution speed, and establishes the basic road right relationship according to the passing priority, to obtain an initial real-time control instruction; Set a risk threshold and an opportunity threshold; based on the risk field and opportunity field data, the vehicle queries the risk field value and opportunity field value on the current position and predicted trajectory in real time; based on the risk field value and opportunity field value, combine the initial real-time control instruction to obtain a real-time vehicle control instruction; Specifically, when the risk field value > risk threshold, it is determined that a high-risk area exists in front of the trajectory, and the basic execution speed in the initial real-time control instruction is directly reduced by a risk value proportion; when the opportunity field value < opportunity threshold, it is determined that a high opportunity area exists near the trajectory, and the basic execution speed in the initial real-time control instruction is increased by an opportunity value proportion; obtain the real-time vehicle control instruction; S43, issue the real-time vehicle control instruction to the vehicle through the cooperative instruction distribution mode to perform cooperative control.

[0024] Further, please refer to Figure 2 A vehicle cooperative decision system based on real-time traffic situation prediction, for realizing the above-mentioned vehicle cooperative decision method based on real-time traffic situation prediction, comprising a data collection and fusion module, a traffic situation prediction module, a risk and opportunity field generation module, and a vehicle control instruction generation and cooperative control module; The data collection and fusion module is used for collecting traffic data from multiple sources in real time and performing integrated processing; by acquiring vehicle data, roadside data and cloud data; all data are aligned to a unified timestamp and coordinate system by using Kalman filtering, to form multi-source traffic fusion data, and a real-time dynamic traffic scene graph is constructed based on this, which includes a static layer and a dynamic layer, and is updated at a fixed period, providing a basis for subsequent analysis; The traffic situation prediction module predicts the traffic situation based on real-time dynamic traffic scene graphs and uses an improved LSTM neural network; by introducing a dynamic graph attention mechanism and a time sequence self-attention mechanism in the graph convolution LSTM unit, the understanding ability of the model for spatio-temporal interaction features is enhanced, so as to construct an initial LSTM traffic situation prediction network; the network is trained using a historical dynamic traffic scene graph with interaction event annotation, the learning rate is optimized by an artificial fish swarm algorithm, and a final prediction network is obtained; the real-time dynamic traffic scene graph is input into the network, and traffic situation prediction data including vehicle future trajectory and interaction event probability are output; The risk and opportunity field generation module converts the traffic situation prediction data into risk field and opportunity field data to quantify the dangers and opportunities in the traffic environment; a target area is divided into a three-dimensional spatio-temporal grid structure, then the conflict probability and uncertainty risk are calculated, and the risk field data are obtained by weighted fusion; the passing gaps in the spatio-temporal grid are identified, the safety margin, stability and practicability are evaluated, the gap quality score is generated, and the opportunity field data are formed by diffusion operation; these field data represent the risk level and opportunity availability in the spatio-temporal dimension, providing a basis for vehicle control; The vehicle control instruction generation and cooperative control module generates real-time vehicle control instructions based on the risk field and opportunity field data through a hierarchical optimization strategy; at the global level, an optimization problem is constructed to minimize the total travel time, a genetic algorithm is used to assign passing priority and recommended speed, and the optimal global instruction is obtained; at the vehicle level, the basic execution speed is adjusted according to the real-time queried risk field and opportunity field values, and the final control instruction is generated; the instructions are distributed to vehicles through a cooperative instruction distribution method to realize cooperative control of multiple vehicles and improve overall traffic efficiency and safety.

[0025] In the description of the present specification, the description of the terms "one embodiment", "example", "specific example" and the like means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the invention. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0026] The preferred embodiments of the above disclosed invention are only used to help explain the invention. The preferred embodiments do not describe all the details and limit the invention to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of the present specification. The present specification selects and describes these embodiments in order to better explain the principles and practical applications of the invention, so that those skilled in the art can well understand and utilize the invention.

Claims

1. A vehicle cooperative decision-making method based on real-time traffic situation prediction, characterized in that, Includes the following steps: S1. Collect and merge vehicle data, roadside data, and cloud data to obtain multi-source traffic fusion data; construct a real-time dynamic traffic scene map based on the multi-source traffic fusion data; S2. Based on the temporal and spatial characteristics of real-time dynamic traffic scene maps, a spatiotemporal graph convolutional network and an attention mechanism are introduced into the LSTM neural network to obtain the initial LSTM traffic situation prediction network. The initial LSTM traffic situation prediction network is trained using a historical dynamic traffic scene map with interactive event annotations to obtain the final LSTM traffic situation prediction network. The real-time dynamic traffic scene map is input into the final LSTM traffic situation prediction network to obtain traffic situation prediction data. S3. Transform traffic situation forecast data into risk field data and opportunity field data; S4. Based on risk field data and opportunity field data, real-time vehicle control commands are obtained through a hierarchical optimization strategy that combines centralized planning with distributed execution. Real-time vehicle control commands are sent to vehicles through collaborative command distribution for collaborative control.

2. The vehicle cooperative decision-making method based on real-time traffic situation prediction according to claim 1, characterized in that, S1 includes the following steps: S11. Collect vehicle data; S12. Collect roadside data; S13, Obtain cloud data; S14. Vehicle data, roadside data, and cloud data are processed by Kalman filtering to obtain multi-source traffic fusion data; S15. Construct a static traffic map layer; construct a real-time dynamic traffic scene map based on multi-source traffic fusion data combined with the static traffic map layer.

3. The vehicle cooperative decision-making method based on real-time traffic situation prediction according to claim 2, characterized in that, S15 includes the following steps: S151. Construct a static traffic map layer; S152. Using the static traffic map layer as the base, dynamic nodes and relational edges are created based on multi-source traffic fusion data to obtain a real-time dynamic traffic scene map.

4. The vehicle cooperative decision-making method based on real-time traffic situation prediction according to claim 1, characterized in that, S2 includes the following steps: S21. Based on the temporal and spatial characteristics of dynamic traffic scene maps, an initial LSTM traffic situation prediction network is obtained by introducing a spatiotemporal graph convolutional network and an attention mechanism into the LSTM neural network and improving the LSTM neural network. S22. Collect historical dynamic traffic scene maps; label key interactive events in the historical dynamic traffic scene maps, and label the start time and vehicles involved in each event to obtain historical dynamic traffic scene maps with interactive events; S23. The initial LSTM traffic situation prediction network is trained using a historical dynamic traffic scene map with interactive events. During the training process, the optimal learning rate of the initial LSTM traffic situation prediction network is found through an optimization algorithm to obtain the optimal solution. The optimal solution is used as the learning rate of the initial LSTM traffic situation prediction network to obtain the final LSTM traffic situation prediction network. S24. Input the real-time dynamic traffic scene map into the final LSTM traffic situation prediction network to obtain traffic situation prediction data.

5. The vehicle cooperative decision-making method based on real-time traffic situation prediction according to claim 4, characterized in that, S21 includes the following steps: S211. Replace the fully connected layers in the input gate, forget gate, and output gate of the LSTM neural network with graph convolutional gate units; S212. A dynamic graph attention mechanism is introduced into the graph convolutional gating unit to calculate the aggregation weights of neighboring nodes. S213. A temporal self-attention mechanism is introduced after the stacked graph convolutional gate units to process the hidden state sequence output by the graph convolutional gate units.

6. The vehicle cooperative decision-making method based on real-time traffic situation prediction according to claim 4, characterized in that, The training process in S23 uses an optimization algorithm to find the optimal learning rate for the initial LSTM traffic situation prediction network, and the optimal solution is obtained by the following steps: S231. Set the prediction accuracy and resource utilization of the initial LSTM traffic situation prediction network; set the maximum number of optimization iterations and the size of the artificial fish swarm's living space; S232. Based on the size of the living space of the artificial fish swarm, construct an initial artificial fish swarm, and set the size of the initial artificial fish swarm according to the learning rate of the initial LSTM traffic situation prediction network. S233. Based on the prediction accuracy and resource utilization rate, define the fitness function of the rate of return brought about by the number of artificial fish swarms; S234. Perform iterative operations on the initial artificial fish population. The higher the fitness value, the better the number of artificial fish populations. In each iteration, calculate the fitness value of the number of artificial fish populations according to the fitness function. Update the number of artificial fish populations according to the fitness value from high to low. In each iteration, obtain the current best number of artificial fish populations and the global best number of artificial fish populations. S235. Repeat S234. When the maximum number of optimization iterations is reached, stop the iteration and take the global optimal fish population as the optimal solution.

7. The vehicle cooperative decision-making method based on real-time traffic situation prediction according to claim 1, characterized in that, S3 includes the following steps: S31. Divide the target area into a two-dimensional grid; assign a unique spatial coordinate to each grid cell in the two-dimensional grid, and establish a mapping relationship with the static elements in the real-time dynamic traffic scene map to obtain a two-dimensional grid with mapping relationship; S32. Divide the prediction period in the traffic situation prediction data into equally spaced time segments to obtain a set of time segments; assign a timestamp to each time segment in the set of time segments to obtain a complete time series from the current moment to the prediction endpoint. The three-dimensional spatiotemporal grid structure is obtained by taking the two-dimensional grid with mapping relationship along the complete time series from the current time to the prediction endpoint; S33. For each pair of interactive vehicles, based on the predicted trajectory distribution in the traffic situation prediction data, calculate the conflict probability at each point in the spatiotemporal grid to obtain the conflict probability data. The covariance matrix of each predicted trajectory is obtained to characterize the prediction uncertainty; multiple trajectories are generated through Monte Carlo sampling, and the probability variance of each spatiotemporal grid cell being occupied by a vehicle is calculated to obtain the uncertainty contribution. The risk field data is obtained by weighting and fusing conflict risk and uncertainty risk, and then smoothly diffusing them in the spatiotemporal dimension through a Gaussian kernel function. S34. Identify continuous areas not occupied by vehicles in the spatiotemporal grid to obtain a set of passage gaps; collect spatiotemporal gap data for each gap in the set of passage gaps; construct a gap quality scoring model based on safety margin, stability, and practicality; process the spatiotemporal gap data through the gap quality scoring model, calculate the quality score for each gap in the set of passage gaps, and obtain a set of gap quality scores. The gap quality scores in the gap quality score set are assigned to the corresponding spatiotemporal grid cells to form the initial opportunity distribution; based on vehicle dynamics constraints, the initial opportunity distribution is subjected to reachability filtering to remove unusable gaps, resulting in the processed opportunity distribution; Spatiotemporal diffusion operations are applied to the high-quality gaps in the processed opportunity distribution to obtain opportunity field data.

8. The vehicle cooperative decision-making method based on real-time traffic situation prediction according to claim 1, characterized in that, S4 includes the following steps: S41. Based on traffic situation prediction data, construct a global optimization problem with the objective of minimizing total travel time; use a genetic algorithm to find the optimal global instruction for the global optimization problem. S42. The global optimal command is issued to each vehicle; each vehicle uses the target speed in the global optimal command as the basic execution speed, and establishes basic right-of-way relationships based on traffic priority to obtain the initial real-time control command. Set risk thresholds and opportunity thresholds; based on risk field and opportunity field data, the vehicle queries its current position and the risk field and opportunity field values ​​on the predicted trajectory in real time; based on the risk field and opportunity field values, combined with the initial real-time control commands, real-time vehicle control commands are obtained. S43. Real-time vehicle control commands are sent to vehicles through collaborative command distribution for collaborative control.

9. A vehicle cooperative decision-making method based on real-time traffic situation prediction according to claim 8, characterized in that, The step of S41, which involves finding the optimal solution to the global optimization problem using a genetic algorithm, includes the following steps: S411. Define the decision variables for the global optimization problem; the decision variables include the travel sequence and the suggested velocity; construct the fitness function for the global optimization problem; S412. Construct an initial chromosome population; each chromosome in the chromosome population represents a set of vehicle passage sequences and suggested speeds; set the maximum number of optimization iterations; S413. Begin iteration. In each iteration, calculate the fitness of each chromosome in the chromosome population according to the second fitness function formula to obtain the chromosome fitness set. Select chromosomes in the chromosome population according to the fitness set to obtain the selected chromosome population. Perform crossover and mutation operations on the chromosomes in the selected chromosome population to obtain the operated chromosome population. S414, repeat S413, and when the maximum number of optimization iterations is reached, the optimal global instruction is obtained.

10. A vehicle cooperative decision-making system based on real-time traffic situation prediction, characterized in that, The system implements a vehicle collaborative decision-making method based on real-time traffic situation prediction as described in any one of claims 1-9, comprising a data collection and fusion module, a traffic situation prediction module, a risk and opportunity field generation module, and a vehicle control command generation and collaborative control module.

Images