Method for forming a group of autonomous vehicles based on trajectory prediction

Abstract

The present application belongs to the field of unmanned driving, and proposes a trajectory prediction-based unmanned vehicle group formation method, which comprises the following steps: step 1. related definitions; step 2. trajectory prediction of unmanned vehicles based on multi-head attention mechanism; and step 3. trajectory prediction-based unmanned vehicle group formation method. The present application provides an unmanned neighbor vehicle set extraction method based on the Pareto optimality theory; proposes a trajectory prediction algorithm for unmanned vehicles based on the multi-head attention mechanism; provides a leading node election algorithm; proposes a trajectory separation degree for measuring the stability of the vehicle group; on this basis, proposes a trajectory prediction-based unmanned vehicle group formation method; and provides evaluation indexes such as average vehicle group trajectory separation degree, node participation rate, communication transmission cost, node survival time, and vehicle group survival time, and constructs a simulation experiment to verify the effectiveness and stability of the proposed unmanned vehicle group.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving, and more specifically to a method for forming an autonomous vehicle swarm based on trajectory prediction. Background Technology

[0002] Vehicle trajectory prediction based on the surrounding environment is one of the key technologies for autonomous driving. It can assist autonomous driving systems in making reasonable planning and decisions, ensuring the safety and driving efficiency of autonomous vehicles in complex traffic environments. Autonomous driving systems rely on sensor devices such as cameras and LiDAR to perceive the environment and capture static and dynamic information of the surrounding environment, providing a rich data source for vehicle trajectory prediction. At present, there are three main types of vehicle trajectory prediction methods: ① physical model-based vehicle trajectory prediction methods; ② deep learning-based vehicle trajectory prediction methods.

[0003] (1) Traditional trajectory prediction methods Traditional trajectory prediction methods mainly rely on sophisticated mathematical and statistical techniques to make predictions based on historical data and predefined models. Xie et al. integrated physics-based and manipulation-based methods to predict vehicle trajectories using interactive multi-models. This method relies on physical laws and kinematic principles to predict the future trajectory of vehicle waves, considering factors such as current position, speed, acceleration, and road constraints to estimate the future path. Batz et al. gave a trajectory prediction method based on mathematical equations, which assumes that the motion of an object can be described by simple mathematical equations, such as constant velocity or constant acceleration models, and estimates the future position based on the current state of the object and its assumed motion dynamics. Lefkopoulos et al. gave a tracking and prediction task based on a Kalman filter. The measurement results from the sensor are combined with the prediction from the mathematical model to estimate the current state of the object and predict its future trajectory. Wang et al. proposed a vehicle trajectory prediction method based on Markov models to capture the probabilistic dependencies between continuous states of an object, using historical data to estimate the transition probability and making predictions based on the most likely state sequence. Reference

[18] proposed a vehicle trajectory prediction method based on game theory and hidden Markov models, which considers the uncertainty in trajectory prediction by representing the future trajectory as a probability distribution. These models utilize statistical techniques to estimate the most probable trajectories and provide confidence metrics. Li et al. proposed a vehicle trajectory prediction method based on Bayesian filters, which combines sensor measurements with dynamic models to estimate the future trajectory of vehicles.

[0004] (2) Trajectory prediction method based on deep learning In recent years, deep learning-based methods for predicting the trajectory of autonomous vehicles have received widespread attention. Deep learning-based methods leverage the power of artificial neural networks to learn complex patterns and relationships from large datasets. He et al. used fully connected layers to extract features and input them into an LSTM to predict vehicle trajectories. Ding et al. used an LSTM encoder to predict the trajectory of autonomous vehicles by encoding the state of the target vehicle. Zyner et al. used a weighted Gaussian Mixture Model (GMM) to predict multimodal trajectories. They used an encoder-decoder model to obtain the parameters of the GMM and clustered the predicted trajectories using the modalities with the highest probabilities. Xing et al. used GMM to identify driving styles and used LSTM and fully connected layers to score sequential data and driving styles to predict vehicle trajectories. Xin et al. used two independent LSTM architectures to predict the target lane and trajectory based on the vehicle's current state and the expected lane. Dai et al. used two sets of LSTM networks to simulate the motion and interaction of nearby vehicles. Ding et al. proposed a set of GRU encoders to characterize paired interactions between vehicles. Tang et al. used an attention mechanism to create a dynamic encoder composed of multiple RNNs sharing parameters to predict multimodal trajectories. Xu et al. introduced a student-teacher network for trajectory prediction, where the student algorithm is based on an LSTM encoder-decoder model and the teacher algorithm is based on a graph convolutional network. Ke et al. proposed DeepTrack, a lightweight vehicle trajectory prediction method based on deep learning, which uses model pruning techniques to reduce the number of parameters and computational complexity of the trajectory prediction model while maintaining good accuracy. Furthermore, Kim et al. proposed a model with a baseline network and trajectory proposal attention, which is a trajectory prediction network that uses a temporal attention mechanism to extract potential multimodal driving information. Yang et al. utilized the spatiotemporal attention mechanism in LSTM networks to perform lane change prediction and trajectory prediction; this method improved prediction accuracy due to the introduction of the spatiotemporal attention mechanism. Kim et al. used multi-head attention to extract lane and vehicle information to predict the future trajectory distribution of vehicles, mainly by using a multi-head attention mechanism to extract information from lanes and vehicles and simulating the influence of each lane on the vehicle itself based on the attention weights extracted from the LSTM encoder.

[0005] In summary, while some research has been conducted on trajectory prediction for autonomous vehicles, the following issues still exist: (1) The historical trajectory between the target vehicle and its neighboring vehicles is not considered: Existing autonomous vehicle trajectory prediction only uses the vehicle's own historical trajectory to predict its future trajectory, but neighboring vehicles can provide the target vehicle with its historical trajectory of environmental perception, which can help the target vehicle predict its future trajectory.

[0006] (2) The future trajectory of vehicles is not considered as the basis for vehicle group modeling: The existing construction of autonomous vehicle groups only considers the motion similarity and communication connectivity between vehicle nodes, without considering vehicle group modeling based on the future trajectory of vehicles. Summary of the Invention

[0007] To address the above problems, this invention proposes a method for extracting neighbor vehicle sets for autonomous vehicles based on Pareto optimality theory; based on this, it proposes a trajectory prediction method for autonomous vehicles based on a multi-head attention mechanism (TPMAM); it also presents a leader node election algorithm; and proposes a trajectory separation degree to measure vehicle group stability. Furthermore, based on these, it proposes a trajectory prediction based Autonomous Vehicle Group Formation method (TPAVGF). Purpose of the invention: To address the shortcomings of current trajectory prediction methods and autonomous vehicle swarms based on trajectory prediction, which fail to consider the historical trajectories of target vehicles and their neighbors, as well as future trajectories, as the basis for swarm modeling, this invention proposes a method for extracting neighbor vehicle sets based on Pareto optimality theory; an autonomous vehicle trajectory prediction algorithm based on a multi-head attention mechanism; a leader node election algorithm; and a trajectory separation degree to measure swarm stability. Based on these, a method for forming an autonomous vehicle swarm based on trajectory prediction is proposed. Evaluation metrics such as average swarm trajectory separation degree, node participation rate, communication transmission cost, node survival time, and swarm survival time are provided. Simulation experiments are conducted to verify the effectiveness and stability of the proposed autonomous vehicle swarm. Selecting autonomous vehicle nodes with similar future trajectories to form a swarm can improve the time for cooperation between vehicle nodes and maintain swarm structure stability. This effectively enables beyond-line-of-sight perception and enhances the ability of autonomous vehicles to understand their surrounding driving environment, providing effective assurance for stable and orderly autonomous driving behavior. This has significant implications and application value for promoting the healthy and rapid development of autonomous driving.

[0008] The technical solution of the present invention specifically includes the following steps: Step 1. Relevant Definitions; Step 2. Trajectory prediction of autonomous vehicles based on multi-head attention mechanism Step 2.1 Extraction of the set of unmanned neighboring vehicles; Step 2.2 Trajectory prediction for unmanned vehicles.

[0009] Step 3: Formation of an autonomous vehicle swarm based on trajectory prediction; Step 3.1 Neighborhood set construction; Step 3.2 Trajectory separation; Step 3.3 Vehicle Group Roles; Step 3.4 Lead node election; Step 3.5 Vehicle group formation algorithm.

[0010] Compared with the prior art, the present invention has the following beneficial effects. The purpose of this invention is to disclose a method for forming an autonomous vehicle swarm based on trajectory prediction, which can effectively achieve beyond-line-of-sight perception and improve the ability of autonomous vehicles to understand the surrounding driving environment. This can effectively help solve problems of single autonomous driving agents such as lane changing, obstacle avoidance, safety distance adjustment, and road traffic efficiency. It has important significance and application value for promoting the healthy and rapid development of autonomous driving in urban scenarios. Attached Figure Description

[0011] Figure 1 Structure diagram of autonomous driving trajectory prediction algorithm based on multi-head attention mechanism Figure 2 Node Neighborhood Set Construction Process Figure 3 Leading the node election process Figure 4 Vehicle group formation algorithm process Figure 5 SUMO simulation experiment scene diagram Figure 6 Comparison of the accuracy of the trajectory prediction method of this invention with RNN, GRU, and LSTM in MAE. Figure 7 Comparison of the trajectory prediction method of this invention with RNN, GRU, and LSTM in terms of MSE accuracy. Figure 8 The trajectory prediction method of this invention is compared with the accuracy of R2 using RNN, GRU, and LSTM. Figure 9 Communication transmission cost Figure 10 Node participation rate Figure 11 Node lifespan Figure 12 Vehicle group survival time Figure 13 Average vehicle group trajectory separation curve at different speeds Figure 14Vehicle group trajectory separation curve over time at maximum vehicle speed of 25 km / h Figure 15 Vehicle group trajectory separation curve over time at maximum vehicle speed of 35 km / h Figure 16 Vehicle group trajectory separation curve over time at maximum vehicle speed of 45 km / h Figure 17 Flowchart of the invention Detailed Implementation The technical solutions provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0012] This invention specifically includes the following three aspects: Step 1. Relevant Definitions; Step 2. Trajectory prediction of autonomous vehicles based on multi-head attention mechanism Step 2.1 Extraction of the set of unmanned neighboring vehicles; Step 2.2 Trajectory prediction for unmanned vehicles.

[0013] Step 3: Formation of an autonomous vehicle swarm based on trajectory prediction; Step 3.1 Neighborhood set construction; Step 3.2 Trajectory separation; Step 3.3 Vehicle Group Roles; Step 3.4 Lead node election; Step 3.5 Vehicle group formation algorithm.

[0014] Step 4. Simulation Experiment and Result Analysis Step 4.1 Simulation Experiment Setup Step 4.2 Evaluation Indicators Step 4.3 Simulation Experiment Results and Analysis Details are as follows: The symbols used in this invention are explained in Table 1.

[0015] Step 1: Relevant Definitions Definition 1: Trajectory sequence Tra (Vehicle Trajectories): represents an autonomous vehicle. The set of trajectories of sampling points over a continuous time period is as follows: in, Represents a node In the sampling coordinate system The lateral and longitudinal displacements at any given moment. Indicates the timestamp of continuous sampling.

[0016] Definition 2 Trajectory Similarity (TSim): Represents the similarity between two autonomous vehicles. and Inter-trajectory sequence and The similarities are as follows: in, Represents trajectory sequence Length, From continuous sampling points Composed of, Represents a node The One sampling point, This represents the distance between two points in a two-dimensional coordinate system, i.e.: in,( and They are and Coordinates in a two-dimensional coordinate system.

[0017] From formulas (2) and (3), it can be seen that the more similar the trajectory sequences between nodes are, the closer the corresponding sampling points are, and the greater the trajectory similarity. .

[0018] Step 2: Autonomous Vehicle Trajectory Prediction Based on Multi-Head Attention Mechanism External interference can affect the future trajectory of autonomous vehicles, thereby impacting the stability of the vehicle corps. To address this issue, we first extract the set of neighboring autonomous vehicles, and then predict the trajectory of autonomous vehicles based on a multi-head attention mechanism.

[0019] (1) Extraction of unmanned neighbor vehicle set 1) Properties of the vehicle set extraction model i) Motion difference driverless vehicles and The motion difference measure measures the differences in motion characteristics between autonomous vehicles, as follows: in, vector is , respectively representing driverless vehicles The velocity, heading angle, and acceleration. From formula (4), we can know... The value is , The larger, and The greater the difference in movement.

[0020] ii) Vehicle activity Vehicle activity level refers to the degree to which an autonomous vehicle frequently changes lanes, taking into account its spatiotemporal characteristics. Higher vehicle activity level indicates more frequent lane changes, as detailed below: in, Indicates vehicle exist Whether a lane change occurs at any given time, and simultaneously assign Weights that change over time , indicating vehicle The more frequently a vehicle changes lanes recently, the higher its activity level.

[0021] iii) Topological compactness driverless vehicles and The topological compactness between them measures exist In spatial grid structure and The degree of spatial correlation between them is as follows: in, Indicates driverless vehicles Is it in the first In each grid topology For the first The weights of each grid topology, For driverless vehicles In the first Area in a grid topology.

[0022] 2) Neighbor vehicle set extraction method i) Neighbor vehicle set extraction model By analyzing the motion differences and topological density between autonomous vehicles, as well as the activity levels of other autonomous vehicles themselves, the target vehicle... A multi-objective optimization model for extracting neighbor vehicle sets can be constructed. for: in, For the target vehicle, For the target vehicle The field of vision, Set the upper limit for neighboring vehicles.

[0023] ii) Solving the neighbor vehicle set extraction model based on multi-objective optimization The method for extracting the neighbor vehicle set of an autonomous driving target vehicle based on multi-objective optimization first transforms the neighbor vehicle set extraction model into an optimization problem with three objectives, and considers the unique properties of the target vehicle itself and the algorithm complexity as constraints. This invention presents a multi-objective evolutionary solution method for the neighbor vehicle set extraction model based on NSGA-II, used to determine the set of neighbor vehicles that have a significant spatiotemporal influence on the target vehicle, as shown in Algorithm 1, and is described in detail below: a) with the target vehicle At the same timestamp The following collection of driverless vehicles The scope is relatively broad, so the preferred method is through the distance function. Determine the range of decision variables for the NSGA-II evolutionary algorithm (lines 1-6).

[0024] b) To The motion difference and topological density of neighboring vehicles within the target vehicle range, as well as their own vehicle activity, are calculated and used as decision variables for the NSGA-II algorithm. (Lines 7-10).

[0025] c) Finally, the Pareto optimal solution set of the model is obtained using the NSGA-II algorithm. That is, for the target vehicle The set of neighboring vehicles that have a significant impact in time and space (lines 11-12).

[0026] (2) Trajectory prediction of autonomous vehicles based on multi-head attention mechanism This invention proposes a trajectory prediction method that integrates historical information and inter-vehicle interaction information. This method extracts historical interaction information between the target vehicle and its neighboring vehicles based on a multi-head attention mechanism interaction feature encoder. Its model structure is as follows: Figure 1 As shown, the overall framework is divided into three parts: historical trajectory extraction module, interactive feature extraction module, and trajectory prediction module.

[0027] 1) Framework for predicting the trajectory of autonomous vehicles i) Historical trajectory extraction module The main function of the historical trajectory encoding module is to perform preliminary feature extraction on the historical trajectory information of autonomous vehicles. The historical trajectory and motion features of autonomous vehicles exhibit a clear temporal sequence. LSTM is used to encode the historical trajectory and motion features of the target autonomous vehicle and its neighboring vehicle sets, obtaining independently encoded contextual information. Each historical timestamp of the target vehicle... The state below After embedding through a fully connected MLP layer, the data is fed into an LSTM encoder for encoding. The final hidden layer state of the encoder is shown below. Overlay as the output of the historical trajectory extraction module : in, MLP is a type of MLP with Multilayer perceptron with weight matrix, The embedding vector is processed by the MLP. Here are the weight and bias matrices for the LSTM.

[0028] ii) Interactive Feature Extraction Module The main function of the interaction feature extraction layer is to extract the spatiotemporal interaction information between the autonomous driving target vehicle and its neighboring vehicles using a multi-head attention mechanism. First, the historical trajectories and motion features of the autonomous driving neighboring vehicles are output through the historical trajectory extraction module. , Then, the output is processed through two convolutional layers and Flattened to obtain a one-dimensional vector. Finally, by summing up all the interest interaction information, we can obtain a tensor. in, Then the output of each attention The attention feature vector is obtained by superimposing the vectors into a one-dimensional vector. The interaction characteristics of the target vehicle were then obtained through GLU. , This is the set of parameters that need to be trained in GLU.

[0029] iii) Trajectory Prediction Module To obtain the vehicle's historical trajectory features and spatiotemporal interaction information, this module will extract the output from the historical trajectory extraction module. With the interaction feature extraction module The concatenation was performed to obtain the fused feature vector. As input for trajectory prediction, LSTM is used to obtain the final trajectory prediction result. The trajectory prediction algorithm uses the fused feature vectors to predict the trajectory of the autonomous vehicle. This problem is a regression problem, and the loss function is as follows: the loss for a single input sample is: in, and They represent the first The difference between the actual coordinates and the predicted coordinates of each timestamp.

[0030] 2) Autonomous vehicle trajectory prediction algorithm The trajectory prediction algorithm for autonomous vehicles is shown in Algorithm 2, and its specific description is as follows: 1) Initialize the weights of the neural network model ; 2) Divide the original training set according to trajectory history and prediction time domain. ,get ; 3) Execute Algorithm 1 to calculate The set of autonomous driving neighbor vehicles for each sample is obtained. ; 4) Based on step size , training set Divide into several equal-sized ; 5) Iterate through each The forward propagation calculation of the prediction model Predicted future trajectories of all samples ; 6) Calculate according to formula (13) Predicted values ​​for all samples Compared with the true value Loss is calculated and the model weights are updated using the backpropagation algorithm. ; (7) Output the trained trajectory prediction model and model weights .

[0031] Step 3: Algorithm for Autonomous Vehicle Swarm Formation Based on Trajectory Prediction (1) Neighborhood set construction When neighboring nodes move in opposite directions or their distances become unstable, the communication link between nodes will be interrupted for a short period of time. In the context of an autonomous vehicle swarm, this means that a node leaves the swarm immediately after joining, making the swarm structure unstable. To reduce the impact of node mobility on the stability of the autonomous vehicle swarm, this invention constructs a set of autonomous vehicles that maintains a stable communication neighborhood within each node.

[0032] By utilizing heading angle and inter-node distance to construct a stable neighborhood set for nodes, autonomous vehicles... neighboring nodes To become [a certain status], the following conditions must be met. Stable neighborhood nodes: in, It is a node and The distance between them; The distance between nodes is represented by a normal distribution. yes The mean is the average distance between nodes throughout the lifecycle of all autonomous vehicles. yes The variance. It is a node and The angle between the velocities. This represents the tolerable range of included angles between nodes. Calculated using cosine similarity, i.e.: in, express exist Time and The difference in horizontal coordinates at time points. express exist Time and The difference in longitudinal coordinates at any given time.

[0033] The process of constructing the neighborhood set of autonomous vehicles is as follows: Figure 2 As shown, the specific construction steps are as follows: 1) All nodes acquire their own motion information through sensors: vehicle ID, previous position, current position, speed, etc.

[0034] 2) Each node's communication range Neighboring nodes periodically broadcast HELLO data packets containing their own motion information.

[0035] 3) When node P receives the HELLO data packet from neighbor node Q, node P calculates the heading angle with node Q using formula (14) and determines whether node Q is a stable neighbor node using formula (15).

[0036] 4) Compare the vehicle ID of node Q with the neighborhood set of node P. If node Q already exists and has been determined to be a stable neighbor node in step 3), update the neighborhood set information of node Q, and simultaneously update the counter. Increment, otherwise The counter is reset to zero; conversely, if node Q does not exist and has been determined to be a stable neighbor node in step 3), then the node Q information is saved, and the counter is reset. Self-incrementing.

[0037] 5) Each node traverses its own neighborhood set The neighboring nodes form a neighborhood set.

[0038] (2) Trajectory separation When the motion characteristics such as speed, position, and trajectory among nodes in an autonomous vehicle swarm differ significantly, some nodes will leave the swarm within a short period, making it difficult for the swarm to maintain stability for an extended period. This invention, based on the future trajectories and motion characteristics of nodes, proposes Trajectory Separation (TSEP) as a stability indicator for the vehicle swarm.

[0039] Based on the trajectory similarity between nodes in formula (1), the global trajectory similarity of the autonomous vehicle swarm is given. : in, This represents the set of leading and coordinating nodes in a vehicle group. Let be a directed edge in the vehicle group, representing a node. Through leading nodes or collaborating nodes Join the car group.

[0040] driverless vehicles and Link connectivity between (Link Connectivity) indicates the tightness of the communication connection between two nodes. The larger, the better and The longer the communication link between them, the more likely it is to be established, specifically as follows: in, This indicates the maximum communication distance between autonomous vehicles. Indicates driverless vehicles and The relative speed. From formula (17), it can be seen that link connectivity is negatively correlated with the speed difference and distance between nodes. The smaller the speed and distance difference between nodes, the longer the communication connection between nodes lasts, and the higher the link connectivity. The larger.

[0041] Based on the link connectivity between nodes Provides a measure of the global link connectivity of the vehicle group. The details are as follows: Self-driving car swarm The trajectory separation can be expressed as: in, (Global Trajectory Similarity) represents the global trajectory similarity of an autonomous vehicle ensemble. (Global Link Connectivity) represents the global link connectivity of an autonomous vehicle swarm. and These are the normalization coefficients for scaling the corresponding dimensions. .

[0042] To verify the effectiveness of trajectory separation in measuring the stability of an autonomous vehicle swarm, the following proof is given: Proposition 1: Trajectory separation is a feasible measure of the stability of an autonomous vehicle swarm.

[0043] Proof: According to formula (19), we can obtain: From formula (20), we can see that about The partial derivative is less than 0, which means the trajectory separation degree is... It will change with the global trajectory similarity of the vehicle group The global trajectory similarity decreases as the trajectories of nodes increase. Therefore, in a swarm of autonomous vehicles, the more similar the future trajectories of nodes, the higher the global trajectory similarity. The higher the value, the lower the trajectory separation and the higher the stability.

[0044] Similarly, it can be concluded that The partial derivative is less than 0, which means the trajectory separation degree is... With link connectivity The global connectivity of the autonomous vehicle swarm decreases as the efficiency increases. The larger the value, the more stable the communication link between nodes, and the better the trajectory separation. The lower the value, the higher the stability of the autonomous vehicle fleet.

[0045] In summary, in an autonomous vehicle swarm, the more similar the trajectories of the nodes and the more stable the communication connections, the less chaotic the swarm becomes, the lower the trajectory separation, and the higher the swarm stability. Therefore, using trajectory separation to measure the stability of an autonomous vehicle swarm is feasible.

[0046] (3) Vehicle Role To improve the stability of autonomous vehicle swarms, this invention employs a master-leading node and auxiliary coordinating nodes to achieve coordination and intelligent interaction among multiple vehicle systems. The vehicle swarm and its members are as follows: 1) Autonomous vehicle swarm G: An autonomous vehicle swarm can be represented as the following quintuple: in, This represents the set of autonomous vehicle nodes within a vehicle group. This represents the set of directed edges that maintain communication connections between nodes within a vehicle group, where the directed edges... Represents a node Through nodes Join the vehicle group and maintain communication connectivity; This represents the set of node roles for autonomous vehicles, including the lead node. Collaborative nodes Follow the node and edge nodes ; A function representing the mapping from nodes to roles within a vehicle group; express There is a set of target nodes for directed edges, namely the set of leading nodes and collaborating nodes.

[0047] 2) Lead Node (LN): When an edge node... The number of neighborhood sets is greater than the threshold , A leader node election can be held when The leading probability function value If it is higher than all edge nodes in its neighborhood set, then Called the leading node, it can be represented as: in, This represents the threshold that the number of neighbors a peripheral node must satisfy in order to be able to elect a leader node.

[0048] The lead node is responsible for communication, computation, and storage management of member information among vehicle group members, thereby maintaining the stability of the vehicle group.

[0049] 3) Cooperative Node (CN): When an edge node... There is a leading node in the neighborhood set. ,and Received leading node The message allowing joining the car group then... The coordinating nodes, referred to as the vehicle group, can be represented as: in, Represents a node Received leading node The message indicated agreement to join the car group.

[0050] The cooperating nodes are responsible for following the addition of nodes and preventing the re-clustering overhead caused by the departure of the leading node, thereby sharing the communication and computing load of the leading node and cooperating to maintain the stability of the vehicle group.

[0051] 4) Follower node (FN): When an autonomous vehicle... There are cooperating nodes in the neighborhood set. ,and Received from the cooperating node The message allowing joining the car group then... A following node in a vehicle group can be represented as: 5) Edge Node (EN): When an edge node... There are no leading or cooperating nodes in the neighborhood set of a given node, and the probability function value of its leading node is [value missing]. If it is not the largest neighborhood set, then This is called an edge node, and it represents the initial state of an autonomous vehicle, which can be expressed as: (4) Leading node election This invention proposes to evaluate the trajectory representation degree of an autonomous vehicle in terms of its overall motion trajectory trend, the motion similarity degree of its motion pattern with that of the autonomous vehicle's neighborhood set, and the link connectivity degree of its communication link continuity with the autonomous vehicle's neighborhood set. Based on this, a probability function is given for the autonomous vehicle to be selected as the leader node.

[0052] 1) Properties related to leader election i) Trajectory characterization driverless vehicles The trajectory characterization degree characterizes The ability to represent the overall trajectory trend of its neighborhood set, namely: in, Represents a node The average trajectory similarity with all nodes in its neighborhood set, i.e.: Property 1: .

[0053] Proof: Property 1 represents driverless vehicles The more similar the trajectory is to the node in the neighborhood set, the better. The greater the trajectory representation, the better.

[0054] According to formulas (26) and (27), we can obtain: Among them, molecules The discriminant of a linear equation in two variables is: From formula (1), the range of values ​​for trajectory similarity between autonomous driving nodes can be obtained as follows: From formulas (29) and (3.30), we can obtain Therefore, the numerator of equation (3.28) has two solutions. ,Right now In this case, according to the quadratic formula, we can obtain: From formulas (29) and (31), we can see that the two solutions... All are less than 0, therefore when hour ,Right now The greater the similarity to the average trajectory of nodes in the neighborhood set, the greater the trajectory representation and the more representative it is of the overall trajectory of the neighborhood set.

[0055] ii) Motion similarity driverless vehicles Motion similarity characterizes The degree of similarity between the movement patterns of its neighboring nodes, i.e.: in, and These are nodes and nodes The difference between velocity and acceleration.

[0056] From formula (32), we can see that, That is, when the velocity and acceleration between nodes are the same, the more similar the trends of motion changes of the nodes, the better. The larger.

[0057] iii) Link connectivity driverless vehicles The link connectivity characterizes The persistence of the communication link between it and its neighboring nodes can be obtained according to formula (16): From formula (33), it can be seen that the link connectivity is negatively correlated with the speed difference and distance between the two nodes. Therefore The smaller the speed and distance differences between the nodes and the neighboring set nodes, the longer the communication connection between the nodes can be maintained. Link connectivity The larger.

[0058] 2) Leader election algorithm driverless vehicles The probability of being elected as the leader node is formally represented using a normalized probability function, which comprehensively considers the autonomous vehicle. The trajectory representation degree, motion similarity, and link connectivity of the relative autonomous driving neighbor list are as follows: in, , and This is the normalization coefficient.

[0059] In summary, in the list of autonomous driving neighbors, autonomous vehicles The more similar the trajectory and movement pattern to the list of autonomous driving neighbors, the greater the link connectivity. Leading probability function The larger the value, The higher the probability of being selected as the leading node.

[0060] The leader node election process based on the leader probability function is as follows: Figure 3 As shown in Algorithm 4, the specific steps are as follows: a) All nodes are initialized as edge nodes and broadcast HELLO packets to their surrounding neighbors.

[0061] b) After each node has the motion information of its neighboring nodes, it calculates its own leading function value according to formula (34). And broadcast the LEADER_PROB data packet.

[0062] c) If the number of neighboring nodes in the neighborhood set of node P reaches the threshold Node P sorts the leading probability function values ​​of its neighboring nodes in the neighborhood set. If node P has the largest leading probability function value, then node P is selected as the leading node; otherwise, it retains its status as an edge node and broadcasts a request to join the vehicle group message JOIN_REQ.

[0063] (5) Vehicle group formation algorithm After the successful construction of the node neighborhood set and the election of the leading node, all nodes possess their own relatively stable neighborhood set information. Based on Algorithm 3 and Algorithm 4, combined with trajectory separation, a complete autonomous vehicle swarm formation algorithm, TPAVGF, is formed. The swarm formation process is as follows: Figure 4 As shown, the details are as follows: 1) All nodes are initialized as edge nodes.

[0064] 2) Each node periodically broadcasts HELLO data packets and receives data packets from neighboring nodes containing motion information such as position and speed. During the broadcasting process, Algorithm 3 is executed to construct its own neighborhood set.

[0065] 3) Traverse each node and execute Algorithm 4 to elect a leader node.

[0066] 4) Edge nodes broadcast a request to join the vehicle group using the REQ_JOIN message, when the leading or coordinating node... After receiving the REQ_JOIN message from node Q, calculate the trajectory separation degree of the vehicle group at this time. When the vehicle group does not reach its maximum size ,like And the current vehicle group trajectory separation Less than the threshold If the condition is met, reply with APROVE_JOIN to agree to join the message; otherwise, reply with DENY_JONE to refuse to join the message.

[0067] 5) If node Q receives the APROVE_JOIN message from node P, then node Q becomes either a collaborating node or a follower node of node P, depending on node P's role; otherwise, node Q continues to broadcast the REQ_JOIN message.

[0068] 6) After an edge node joins the vehicle group, the leading node and the cooperating node maintain the connection by periodically sending HEARTBEAT heartbeat packets to their subordinate nodes. Similarly, the cooperating nodes and the following nodes send HEARTBEAT heartbeat packets to their superior nodes to maintain the connection.

[0069] The specific algorithm for forming a swarm of autonomous vehicles is shown below: Step 4. Simulation Experiment and Result Analysis To verify the effectiveness of the autonomous vehicle trajectory prediction algorithm TPMAM and the autonomous vehicle swarm formation method TPAVGF, this invention constructs a simulation experimental scenario using SUMO and conducts simulation experiments.

[0070] (1) Simulation and experimental setup 1) Experimental setup for trajectory prediction method i) The experimental environment used for trajectory prediction of unmanned vehicles is shown in Table 2.

[0071] ii) Dataset The experiment used the Next Generation Simulation (NGSIM) dataset for trajectory prediction of autonomous vehicles. This dataset, collected and labeled by the Federal Highway Administration (FHMA), provides detailed vehicle trajectory information, such as vehicle location coordinates and speed, for highways in multiple regions of the United States. Each dataset provides 45 minutes of traffic data for a road segment at a sampling frequency of 10 Hz, and includes 15 minutes of road segment data under light, moderate, and congested traffic conditions. This chapter uses the datasets for two road segments, US-101 and I-80, as detailed in Table 3.

[0072] 2) Experimental setup for driverless vehicle swarms The simulation experiments used SUMO (Simulation of Urban Mobility) to simulate vehicle motion and road topology, and NS3 (Network Simulator 3) to simulate interactions between vehicles. SUMO is a microscopic urban traffic flow simulation software that reproduces the operation of the road network environment by simulating facilities such as vehicles, road networks, and traffic lights. NS3 is a discrete event network simulator written in C++ that allows users to simulate the behavior of vehicle nodes in the network: starting position, direction, speed, and trajectory, etc. This invention uses SUMO to simulate traffic flow data of autonomous vehicles in a real road network environment, and then uses NS3 to receive vehicle driving data exported from SUMO to simulate network communication between autonomous vehicles.

[0073] like Figure 5 As shown, this invention uses SUMO to construct a 2km long and 1km wide grid road, and deploys 200 vehicles for simulation. Considering that autonomous vehicles need historical trajectory data for trajectory prediction, the simulation start time is set after 80 seconds, and 200 seconds of the simulation process is extracted for evaluation and analysis of experimental indicators. To analyze the performance of the vehicle swarm formation algorithm under various speed scenarios, the maximum speed of the autonomous vehicles is set in the range of 25km / h to 50km / h, increasing in increments of 5km / h. In NS3, short-range communication between vehicle nodes is set via the WAVE protocol, with a maximum communication range of 250m. Throughout the simulation experiment, motion information such as the position and speed of the autonomous vehicles and network communication data packets are collected every second. Specific simulation experimental parameter settings are shown in Table 4.

[0074] To verify the effectiveness of the autonomous vehicle swarm formation algorithm TPAVGF, the following simulation experiment was designed: (1) Use SUMO to construct a simulation test scenario including unmanned vehicles. Configure relevant simulation parameters and basic attributes of unmanned vehicles according to Table 1. Simulate and export vehicle trajectory and behavior data under different maximum speed scenarios (25km / h-50km / h) of unmanned vehicles.

[0075] (2) Import vehicle trajectory and behavior data into NS3, implement the vehicle group formation algorithm TPAVGF, run simulation experiments in NS3 and collect data of unmanned vehicles in different speed scenarios.

[0076] (3) Collect the position, speed and network communication behavior of the unmanned vehicle during the simulation experiment, and analyze the effectiveness of the TPAVGF algorithm.

[0077] (4) The TPAVGF algorithm is compared with the SCCAVGF algorithm proposed by JJ Cheng et al. Based on the above five evaluation indicators, the performance of the TPAVGF algorithm is analyzed.

[0078] (2) Evaluation indicators 1) Evaluation indicators for trajectory prediction of autonomous vehicles To evaluate the performance of the autonomous vehicle trajectory prediction algorithm based on the multi-head attention mechanism proposed in this invention, the proposed TPDL algorithm was compared with the Gated Recurrent Unit (GRU) model, Recurrent Neural Network (RNN) model, and Long Short-Term Memory (LSTM) model. The specific evaluation metrics are as follows: (1) Mean Squared Error (MSE): This represents the mean of the sum of squares of the differences between the actual trajectory values ​​and the predicted trajectory values ​​of the autonomous vehicle. The smaller the MSE, the better the model's predictive ability. Specifically: in, Represents the total number of samples. Represents the time domain of trajectory prediction. Indicates the first One sample in The actual trajectory value at any given moment. This represents the predicted trajectory value for the corresponding sample.

[0079] (2) Mean Absolute Error (MAE): This represents the mean of the sum of the absolute values ​​of the differences between the actual trajectory values ​​and the predicted trajectory values ​​of the autonomous vehicle. Since MAE calculates the absolute value of the error, outliers have a smaller impact on the overall mean. Specifically: (3) R² coefficient of determination: indicates the degree of fit of the model to the real trajectory data, R²∈[0,1]. The larger the R², the better the fit, as detailed below: in, Indicates the first Sample The mean of the true trajectory values ​​in the time domain.

[0080] 2) Evaluation Indicators for Autonomous Vehicle Flock The following five metrics were set to evaluate the performance of the autonomous vehicle swarm formation algorithm in terms of the effectiveness and stability of vehicle swarm formation.

[0081] (1) Average vehicle group trajectory separation: Vehicle group trajectory separation As defined by formula (19), the vehicle group trajectory separation degree represents the average trajectory separation degree of all vehicles in the group per unit time during their life cycle in the simulation experiment. As can be seen from Proposition 1, this index can evaluate the overall stability of the autonomous vehicle group.

[0082] in, Indicates the average vehicle group trajectory separation. Indicates the number of autonomous vehicles in the fleet. Indicates a group of vehicles The life cycle.

[0083] (2) Node participation rate: This represents the proportion of vehicle nodes in the autonomous vehicle swarm to the total number of nodes per unit time during the simulation experiment. The higher the node participation rate, the more nodes successfully join the swarm, and the better the effectiveness of the swarm formation method.

[0084] (3) Communication transmission cost: This refers to the total amount of data transmitted per unit time by all autonomous vehicles participating in the formation of the vehicle swarm during the entire simulation experiment. The more frequent the communication between nodes and the larger the data packet capacity, the higher the data transmission cost, resulting in greater overhead for vehicle swarm formation, as detailed below: in, This represents the total amount of data transmitted through communication between all nodes during the entire simulation experiment. This indicates the size of each data packet. Indicates the simulation time.

[0085] (4) Node survival time: This refers to the time that the unmanned vehicle is in the vehicle group during the entire simulation experiment. It is also the time that the node remains in the non-edge node state. The longer the node survival time, the higher the efficiency of the node joining the vehicle group and the better the effectiveness of the vehicle group formation method.

[0086] (5) Vehicle swarm survival time: This refers to the life cycle of the autonomous vehicle swarm, that is, the cumulative time from the formation to the extinction of the autonomous vehicle swarm. The longer the vehicle swarm survival time, the better the overall stability of the vehicle swarm.

[0087] (3) Analysis of simulation and experimental results 1) Accuracy of autonomous vehicle trajectory prediction i) Comparison of the accuracy of the trajectory prediction method of this invention with RNN, GRU, and LSTM in MAE. Figure 6 This is a comparison chart of the MAE (Modular Value Exceeded) of RNN, GRU, LSTM, and the method of this invention. MAE represents the average of the absolute differences between the predicted trajectory values ​​of the autonomous vehicle and the actual trajectory values ​​of the four methods. The lower the MAE value, the more accurate the autonomous vehicle trajectory prediction method. Figure 6 As can be seen, the MAE values ​​of the four methods gradually increase with the increase of the prediction time domain. This is because trajectory prediction requires the trajectory predicted at the previous time moment as input. As the error accumulates, the accuracy of trajectory prediction decreases accordingly for longer time domains. However, the MAE values ​​of the algorithm of this invention within the next 1-5 seconds are higher than those of the comparison algorithms. The figure shows that the MAE values ​​of the method of this invention within the next 1-5 seconds are 0.0133, 0.0152, 0.016, 0.0162, and 0.017, respectively, with an average of 0.01565. The average values ​​of the RNN, GRU, and LSTM algorithms within the next 1-5 seconds are 0.06932, 0.0458, and 0.03388, respectively. The average future time MAE values ​​of the algorithm of this invention are 0.05367, 0.03015, and 0.01823 lower than those of the comparison methods, indicating that the method of this invention has higher accuracy than trajectory prediction methods based on RNN, GRU, and LSTM.

[0088] ii) Comparison of the accuracy of the trajectory prediction method of this invention with RNN, GRU, and LSTM in MSE. Figure 7 This is a comparison chart of the MSE (Mean Squared Error) of the method of this invention and the comparative methods. MSE represents the mean of the sum of squared differences between the predicted trajectory values ​​of the autonomous vehicle and the actual trajectory values ​​of the four methods. The smaller the MSE, the better the model's predictive ability. Figure 7As can be seen, the MSE value of the algorithm of this invention within 1-5 seconds of the future time is lower than that of the comparative algorithms. Specifically, the average MAE value of the method of this invention within 1-5 seconds of the future time is 0.00045225, which is significantly higher than the average value of the algorithms based on RNN, GRU, and LSTM within 1-5 seconds of the future time. This indicates that the method of this invention has a higher accuracy than future trajectory prediction methods based on RNN, GRU, and LSTM.

[0089] ii) Comparison of the accuracy of the trajectory prediction method of this invention with RNN, GRU, and LSTM in MSE. Figure 8 This is an R² comparison graph of the method of this invention and the comparative methods. R² represents the degree of fit between the trajectories predicted by the four methods and the real trajectory data. The larger the R², the better the fit. From Figure 8 As can be seen, the R2 value of the algorithm of the present invention is higher than that of the comparison algorithm within 1-5 seconds in the future. This indicates that the method of the present invention fits the real trajectory better than the trajectory prediction based on RNN, GRU and LSTM. In other words, the future trajectory predicted by the method of the present invention is more accurate.

[0090] 2) Analysis of Experimental Results of Autonomous Vehicle Swarm To meet the requirements of autonomous vehicles in various speed scenarios, the maximum speed of autonomous vehicles was gradually increased from 25 km / h to 50 km / h in 5 km / h increments, selecting a total of six representative speed scenarios. Simulation experiments were conducted in each speed scenario, and the node participation rate, communication transmission cost, node survival time, vehicle group survival time, and vehicle group trajectory separation degree of the vehicle group were recorded. This was used to evaluate the performance of the TPAVGF and SCCAVGF algorithms in terms of the effectiveness of autonomous vehicle group formation and vehicle group stability.

[0091] i) Effectiveness of autonomous vehicle swarm formation methods Communication transmission cost reflects the total amount of data transmitted by all autonomous vehicles per unit time. Lower communication transmission cost indicates lower communication overhead for vehicle nodes during vehicle swarm formation, and a more efficient vehicle swarm formation algorithm. For example... Figure 9As shown, the red line represents the TPAVGF algorithm, whose communication transmission cost reaches a maximum of 2.67 KB / s when the maximum speed of the autonomous vehicle swarm is 35 km / h, and remains around 2.65 KB / s at other speeds. The green line represents the SCCAVGF algorithm, whose communication transmission cost reaches a minimum of 9.43 KB / s when the maximum speed of the autonomous vehicles is 45 km / h, and a maximum of 11.27 KB / s when the maximum speed of the autonomous vehicles is 35 km / h, showing a large fluctuation range. Simulation results show that the TPAVGF algorithm's communication transmission cost is significantly lower than that of the SCCAVGF algorithm across different speed scenarios. The main reason is that the SCCAVGF algorithm uses a distributed vehicle swarm architecture, where all nodes within the swarm need to communicate with each other during swarm formation, greatly increasing communication overhead. In contrast, the TPAVGF algorithm uses a semi-centralized vehicle swarm architecture, where all nodes only need to communicate with their own superior and subordinate nodes, significantly reducing communication transmission costs. In conclusion, during vehicle platoon formation, the TPAVGF algorithm has lower communication overhead than the SCCAVGF algorithm and is applicable to different speed scenarios, indicating that the TPAVGF algorithm is more effective.

[0092] Node participation rate reflects the proportion of nodes with all autonomous vehicles in the foyer to the total number of nodes per unit time. A higher node participation rate indicates that the foyer covers more autonomous vehicles, and the foyer formation algorithm is more effective. Figure 10 As shown in the figure, the node participation rate of the TPAVGF algorithm (represented by the red line) shows a generally stable upward trend with the increase of the maximum speed of the autonomous vehicle. It reaches a minimum of 0.837 and a maximum of 0.86 at maximum speeds of 30 km / h and 40 km / h, respectively, and remains around 0.85 overall. The node participation rate of the SCCAVGF algorithm (represented by the green line) reaches a minimum of 0.683 at a maximum speed of 30 km / h, and remains around 0.726 at other speeds. Simulation results show that the node participation rate of the TPAVGF algorithm is significantly higher than that of the SCCAVGF algorithm across different speed scenarios. This is mainly because the TPAVGF algorithm adopts a semi-centralized vehicle group architecture with leading nodes as the main component and collaborating nodes as secondary components, expanding the communication range of the vehicle group through collaborating nodes. In conclusion, the TPAVGF algorithm achieves a higher node participation rate across different speed scenarios, indicating that the TPAVGF algorithm is more effective.

[0093] Node lifetime reflects the time a node spends within the vehicle horde during its lifecycle. A longer lifetime indicates higher efficiency in node joining the horde and better effectiveness of the horde formation algorithm. For example... Figure 11As shown, the node survival time of the TPAVGF algorithm (represented by the red line) remains around 163.8 seconds as the maximum speed of the autonomous vehicle increases, exhibiting only minor fluctuations. The node survival time of the SCCAVGF algorithm (represented by the green line) reaches a minimum of 134 seconds at a maximum speed of 30 km / h, and remains around 142.2 seconds at other speeds. In summary, the node survival time of the TPAVGF algorithm is significantly longer than that of the SCCAVGF algorithm across different speed scenarios. Furthermore, from... Figure 11 It can be seen that as the maximum speed of autonomous vehicles increases, the changing trends of node survival time and node participation rate are consistent. This is because both indicators are statistical analyses of the state of non-edge nodes in the vehicle swarm.

[0094] i) Stability of autonomous vehicle swarms The survival time of the autonomous vehicle swarm reflects the average survival time of the autonomous vehicle swarm throughout the entire simulation experiment. A higher vehicle swarm survival time indicates that the autonomous vehicle swarm can survive for a longer time under external interference and that the vehicle swarm is more stable.

[0095] Figure 12 This paper illustrates the relationship between the TPAVGF and SCCAVGF algorithms and the maximum speed of autonomous vehicles, where the red line represents the TPAVGF algorithm and the green line represents the SCCAVGF algorithm. From... Figure 12 It can be seen that when the maximum speed of autonomous vehicles is between 30km / h and 40km / h, the vehicle swarm survival times of the TPAVGF and SCCAVGF algorithms are not significantly different, both maintaining above 31s. However, at maximum speeds of 25km / h, 45km / h, and 50km / h, the vehicle swarm survival times of the TPAVGF algorithm are 36.5s, 38s, and 33.3s respectively, significantly higher than the SCCAVGF algorithm, indicating that the TPAVGF algorithm performs better in low- and high-speed scenarios. This is because the TPAVGF algorithm considers the future trajectories of vehicles as the basis for vehicle swarm modeling, while the SCCAVGF algorithm only considers the motion similarity between vehicle nodes. Vehicles formed by the TPAVGF algorithm have more similar future trajectories than those formed by the SCCAVGF algorithm, allowing them to maintain links for a longer period and resulting in higher swarm stability. In conclusion, the TPAVGF algorithm can maintain high vehicle swarm stability in most speed scenarios.

[0096] The average vehicle group trajectory separation reflects the overall stability of the vehicle group. A smaller average vehicle group trajectory separation indicates that the trajectories of the leading and cooperating nodes are more similar to those of their subordinate nodes, the communication link is more stable, and the overall stability of the vehicle group is higher. In this simulation experiment, the maximum speed of the autonomous vehicles is set within the range of 25 km / h to 50 km / h, increasing in 5 km / h increments. Six motion scenarios covering different speeds are selected, and the vehicle group trajectory separation threshold is set accordingly. Simulation experiments were conducted on the TPAVGF and SCCAVGF vehicle group formation algorithms with values ​​of 20, 40, and 60 respectively, and the average vehicle group trajectory separation degree of the autonomous vehicle group was statistically analyzed.

[0097] Figure 13 This chart shows the relationship between the average vehicle group trajectory separation and the maximum speed of the autonomous vehicles for the SCCAVGF and TPAVGF algorithms under different trajectory separation thresholds. The green, purple, and blue lines represent the TPAVGF algorithm at trajectory separation thresholds of 20, 40, and 60, respectively, while the red line represents the SCCAVGF algorithm. Figure 13 It can be seen that the average vehicle group trajectory separation of the TPAVGF algorithm remains around 31.8 under different trajectory separation thresholds, reaching a maximum of 35.7 when the trajectory separation threshold is 20 and the maximum speed of the autonomous vehicles is 45 km / h. In contrast, the average vehicle group trajectory separation of the SCCAVGF algorithm fluctuates significantly with changes in the maximum speed of the autonomous vehicles, reaching a minimum of 43.74 when the maximum speed is 40 km / h, far lower than the worst-case scenario of the TPAVGF algorithm. The TPAVGF algorithm exhibits different performance at different trajectory separation thresholds; in most speed scenarios, it performs best when the trajectory separation threshold is 60, while its performance is relatively poor when the threshold is 20.

[0098] Within the maximum speed range of 25km / h to 50km / h for autonomous vehicles, three representative speed scenarios of 25km / h, 35km / h, and 45km / h were selected. Based on the above results, the trajectory separation threshold of the TPAVGF algorithm was set to 60, and simulation experiments were conducted with the SCCAVGF algorithm. The average trajectory separation of the vehicle group during the entire 200s simulation process was statistically analyzed.

[0099] like Figure 14As shown, when the maximum speed of the autonomous vehicles is 25 km / h, in the initial 121-139 s interval of the simulation experiment, the vehicle trajectory separation degree of the TPAVGF algorithm (represented by the red line) is slightly lower than that of the SCCAVGF algorithm (represented by the green line). During other simulation periods, the vehicle trajectory separation degree of the TPAVGF algorithm is significantly lower than that of the SCCAVGF algorithm. The simulation results demonstrate that the TPAVGF algorithm can achieve high vehicle group stability in low-speed scenarios.

[0100] like Figure 15 As shown, when the maximum speed of the autonomous vehicle is 35 km / h, in the initial 121s-145s interval of the simulation experiment, the vehicle trajectory separation degree of the TPAVGF algorithm (represented by the red line) is not significantly different from that of the SCCAVGF algorithm (represented by the green line). During other simulation periods, the vehicle trajectory separation degree of the TPAVGF algorithm is significantly lower than that of the SCCAVGF algorithm, and both show a steady trend of change with increasing simulation time. The simulation results indicate that the TPAVGF algorithm can achieve high vehicle group stability in medium-speed scenarios.

[0101] like Figure 16 As shown, when the maximum speed of the autonomous vehicles is 45 km / h, the overall vehicle trajectory separation degree of the TPAVGF algorithm (represented by the red line) is lower than that of the SCCAVGF algorithm (represented by the green line) throughout the simulation time. Simulation results indicate that the TPAVGF algorithm exhibits better vehicle group stability in high-speed scenarios.

[0102] In summary, under different speed scenarios, the TPAVGF algorithm achieves low vehicle trajectory separation and higher overall vehicle group stability.

[0103] Innovation This invention addresses the shortcomings of current trajectory prediction methods and autonomous vehicle swarms based on trajectory prediction, which fail to consider the historical trajectories of target vehicles and their neighbors, as well as future trajectories, as the basis for swarm modeling. It proposes a method for extracting neighbor vehicle sets based on Pareto optimality theory; an autonomous vehicle trajectory prediction algorithm based on multi-head attention mechanism; a leader node election algorithm; and a trajectory separation degree to measure swarm stability. Based on these, a method for forming an autonomous vehicle swarm based on trajectory prediction is proposed. Evaluation metrics such as average swarm trajectory separation degree, node participation rate, communication transmission cost, node survival time, and swarm survival time are provided. Simulation experiments are conducted to verify the effectiveness and stability of the proposed autonomous vehicle swarm. Selecting autonomous vehicle nodes with similar future trajectories to form a swarm can improve the time for cooperation between vehicle nodes and maintain swarm structure stability. This effectively enables beyond-line-of-sight perception and enhances the ability of autonomous vehicles to understand their surrounding driving environment. Therefore, it provides effective safety assurance for autonomous vehicle swarms to effectively solve problems such as lane changing, obstacle avoidance, safe distance adjustment, and road traffic efficiency, which is of great significance and application value for promoting the healthy and rapid development of autonomous driving.

[0104] Table 1. Symbol Explanation Table 2. Trajectory Prediction Experiment Parameter Settings Table 3. Description of relevant fields in the NGSIM dataset Table 4 Simulation Experiment Parameter Settings .

Claims

1. A method for forming an autonomous vehicle swarm based on trajectory prediction, characterized in that, Includes the following steps: Step 1. Relevant Definitions; Step 2. Autonomous vehicle trajectory prediction based on multi-head attention mechanism; including: Step 2.1 Extraction of the set of unmanned neighboring vehicles; Step 2.2 Unmanned vehicle trajectory prediction; Step 3: Formation of an autonomous vehicle swarm based on trajectory prediction; including: Step 3.1 Neighborhood set construction; Step 3.2 Trajectory Separation; Step 3.3 Vehicle Group Roles; Step 3.4 Lead node election; Step 3.5 Vehicle group formation algorithm; specifically: Based on the node neighborhood set construction algorithm and the leader node election algorithm, referred to as Algorithm 3 and Algorithm 4 respectively, combined with trajectory separation degree, a complete autonomous vehicle swarm formation algorithm TPAVGF is formed, as follows: (1) All nodes are initialized as edge nodes; (2) Each node periodically broadcasts HELLO data packets and receives data packets from neighboring nodes containing motion information such as position and speed. During the broadcasting process, Algorithm 3 is executed to construct its own neighborhood set. (3) Traverse each node and execute Algorithm 4 to elect the leader node; (4) Edge nodes broadcast a request to join the vehicle group using the REQ_JOIN message, when the leading or coordinating node... After receiving the REQ_JOIN message from node Q, calculate the trajectory separation degree of the vehicle group at this time. When the vehicle group does not reach its maximum size ,like And the current vehicle group trajectory separation Less than the threshold If yes, reply with APROVE_JOIN to agree to join; otherwise, reply with DENY_JONE to refuse to join. (5) If node Q receives the APROVE_JOIN message from node P, then node Q becomes either a collaborating node or a follower node of node P, depending on node P's role; otherwise, node Q continues to broadcast the REQ_JOIN message. (6) After the edge node joins the vehicle group, the leading node and the cooperating node maintain the connection by periodically sending HEARTBEAT heartbeat packets to their subordinate nodes point-to-point. Similarly, the cooperating node and the following node send HEARTBEAT heartbeat packets to their superior nodes to maintain the connection. The specific algorithm for forming a swarm of autonomous vehicles is shown below:

2. The method for forming an autonomous vehicle swarm based on trajectory prediction as described in claim 1, characterized in that, Step 1: Definition 1: Trajectory sequence Tra (Vehicle Trajectories): represents an autonomous vehicle. The set of trajectories of sampling points over a continuous time period is as follows: in, Represents a node In the sampling coordinate system The lateral and longitudinal displacements at any given moment. Indicates the timestamp of continuous sampling; Definition 2 Trajectory Similarity (TSim): Represents the similarity between two autonomous vehicles. and Inter-trajectory sequence and The similarities are as follows: in, Represents trajectory sequence Length, From continuous sampling points Composed of, Represents a node The One sampling point, This represents the distance between two points in a two-dimensional coordinate system, i.e.: in,( and They are and Coordinates in a two-dimensional coordinate system; From formulas (2) and (3), it can be seen that the more similar the trajectory sequences between nodes are, the closer the corresponding sampling points are, and the greater the trajectory similarity. .

3. The method for forming an autonomous vehicle swarm based on trajectory prediction as described in claim 1, characterized in that, Step 2: Step 2.1 Extraction of the set of unmanned neighboring vehicles (1) Properties of the vehicle set extraction model 1) Motion difference driverless vehicles and The motion difference measure measures the differences in motion characteristics between autonomous vehicles, as follows: in, vector is , respectively representing driverless vehicles The velocity, heading angle, and acceleration; from formula (4) we know The value is , The larger, and The greater the difference in movement; 2) Vehicle activity level Vehicle activity level refers to the degree to which an autonomous vehicle frequently changes lanes, taking into account its spatiotemporal characteristics. Higher vehicle activity level indicates more frequent lane changes, as detailed below: in, Indicates vehicle exist Whether a lane change occurs at any given time, and simultaneously assign Weights that change over time , indicating vehicle The more frequently a vehicle changes lanes recently, the higher its activity level. 3) Topological compactness driverless vehicles and The topological compactness between them measures exist In spatial grid structure and The degree of spatial correlation between them is as follows: in, Indicates driverless vehicles Is it in the first In each grid topology For the first The weights of each grid topology, For driverless vehicles In the first Area in a grid topology; (2) Neighbor vehicle set extraction method 1) Neighbor vehicle set extraction model By analyzing the motion differences and topological density between autonomous vehicles, as well as the activity levels of other autonomous vehicles themselves, the target vehicle... Construct a multi-objective optimization model for extracting neighbor vehicle sets. for: in, For the target vehicle, For the target vehicle The field of vision, Set the maximum number of vehicles that can be grouped together from neighboring vehicles; 2) Solving the neighbor vehicle set extraction model based on multi-objective optimization A method for extracting the neighbor vehicle set of an autonomous driving target vehicle based on multi-objective optimization is proposed. First, the neighbor vehicle set extraction model is transformed into an optimization problem with three objectives, considering the unique properties of the target vehicle and the algorithm complexity as constraints. A multi-objective evolutionary solution method based on NSGA-II is given to determine the neighbor vehicle set that has a significant spatiotemporal impact on the target vehicle, as shown in Algorithm 1, and is described in detail below: 1) With the target vehicle At the same timestamp The following collection of driverless vehicles The scope is relatively broad, so the preferred method is through the distance function. Determine the range of decision variables for the NSGA-II evolutionary algorithm; 2) To The motion difference and topological density of neighboring vehicles within the target vehicle range, as well as their own vehicle activity, are calculated and used as decision variables for the NSGA-II algorithm. ; 3) Finally, the Pareto optimal solution set of the model is obtained using the NSGA-II algorithm. That is, for the target vehicle A set of neighboring vehicles that have a significant impact in time and space; Step 2.2 Unmanned vehicle trajectory prediction (1) Framework for predicting the trajectory of unmanned vehicles 1) Historical trajectory extraction module The main function of the historical trajectory encoding module is to perform preliminary feature extraction on the historical trajectory information of autonomous vehicles. The historical trajectory and motion features of autonomous vehicles exhibit a clear temporal sequence. LSTM is used to encode the historical trajectory and motion features of the target autonomous vehicle and its neighboring vehicle sets to obtain independently encoded contextual information. Each historical timestamp of the target vehicle... The state below After embedding through a fully connected MLP layer, the data is fed into an LSTM encoder for encoding. The final hidden layer state of the LSTM encoder is shown below. Overlay as the output of the historical trajectory extraction module : in, MLP is a type of MLP with Multilayer perceptron with weight matrix, The embedding vector is processed by the MLP. Here are the weight and bias matrices for the LSTM; 2) Interaction Feature Extraction Module The interaction feature extraction layer uses a multi-head attention mechanism to extract the spatiotemporal interaction information between the autonomous driving target vehicle and its neighboring vehicles. First, the historical trajectories and motion features of the autonomous driving neighboring vehicles are output through the historical trajectory extraction module. , Then, the output is processed through two convolutional layers and Flattened to obtain a one-dimensional vector. Finally, by summing all the interest interaction information together, we obtain a tensor. in, Then the output of each attention The attention feature vector is obtained by superimposing the vectors into a one-dimensional vector. The interaction characteristics of the target vehicle were then obtained through GLU. , This is the set of parameters that need to be trained in GLU; 2) Trajectory Prediction Module To obtain the vehicle's historical trajectory features and spatiotemporal interaction information, this module will extract the output from the historical trajectory extraction module. With the interaction feature extraction module The concatenation was performed to obtain the fused feature vector. As input for trajectory prediction, LSTM is used to obtain the final trajectory prediction result; the trajectory prediction algorithm uses the fused feature vector to predict the trajectory of the autonomous vehicle; this problem is a regression problem, and the loss function is as follows, with the loss for a single input sample being: in, and They represent the first The difference between the actual coordinates and the predicted coordinates of each timestamp; (2) Autonomous vehicle trajectory prediction algorithm The trajectory prediction algorithm for autonomous vehicles is shown in Algorithm 2, and its specific description is as follows: 1) Initialize the weights of the neural network model ; 2) Divide the original training set according to trajectory history and prediction time domain. ,get ; 3) Execute Algorithm 1 to calculate The set of autonomous driving neighbor vehicles for each sample is obtained. ; 4) Based on step size , training set Divide into several equal-sized ; 5) Iterate through each The forward propagation calculation of the prediction model Predicted future trajectories of all samples ; 6) Calculate according to formula (13) Predicted values ​​for all samples Compared with the true value Loss is calculated and the model weights are updated using the backpropagation algorithm. ; 7) Output the trained trajectory prediction model and its weights. ; 4. The method for forming an autonomous vehicle swarm based on trajectory prediction as described in claim 1, characterized in that, Step 3: Step 3.1 Domain Set Construction When neighboring nodes move in opposite directions or their distances are unstable, the communication link between nodes will be interrupted for a short period of time. In the context of an autonomous vehicle swarm, this means that a node leaves the swarm immediately after joining it, making the swarm structure unstable. To reduce the impact of node mobility on the stability of the autonomous vehicle swarm, a set of autonomous vehicles that remains stable within the node's communication neighborhood is constructed. By utilizing heading angle and inter-node distance to construct a stable neighborhood set for nodes, autonomous vehicles... neighboring nodes To become [a certain status], the following conditions must be met. Stable neighborhood nodes: in, It is a node and The distance between them; The distance between nodes is represented by a normal distribution. yes The mean is the average distance between nodes throughout the lifecycle of all autonomous vehicles. yes The variance; It is a node and The angle between the velocities. This represents the tolerable range of included angles between nodes. Calculated using cosine similarity, i.e.: in, express exist Time and The difference in horizontal coordinates at time points. express exist Time and The difference in vertical coordinates at any given time; The specific steps for constructing the neighborhood set of autonomous vehicles are as follows: (1) All nodes acquire their own motion information through sensors: vehicle ID, previous position, current position, and speed; (2) Each node to the communication range Neighboring nodes within the node periodically broadcast HELLO data packets containing their respective motion information; (3) When node P receives the HELLO data packet from neighbor node Q, node P calculates the heading angle with node Q using formula (15) and determines whether node Q is a stable neighbor node using formula (14). (4) Compare the vehicle ID of node Q with the neighborhood set of node P. If node Q already exists and has been determined to be a stable neighbor node in step (3), then update the information of node Q in the neighborhood set, and simultaneously update the counter. Increment, otherwise The counter is reset to zero; conversely, if node Q does not exist and is determined to be a stable neighbor node through step (3), then the node Q information is saved, and the counter is reset. Incrementing; (5) Each node traverses its own neighborhood set The neighboring nodes form a neighborhood set; Step 3.2 Trajectory Separation Based on the future trajectory and motion characteristics of nodes, the Trajectory Separation (TSEP) is given as a stability index for the vehicle group. Based on the trajectory similarity between nodes in formula (1), the global trajectory similarity of the autonomous vehicle swarm is given. : in, This represents the set of leading and coordinating nodes in a vehicle group. Let be a directed edge in the vehicle group, representing a node. Through leading nodes or collaborating nodes Join the car group; driverless vehicles and Link connectivity between (Link Connectivity) indicates the tightness of the communication connection between two nodes. The larger, the better and The longer the communication link between them, the more likely it is to be established, specifically as follows: in, This indicates the maximum communication distance between autonomous vehicles. Indicates driverless vehicles and The relative speed; as shown in formula (17), the link connectivity is negatively correlated with the speed difference and distance between nodes. The smaller the speed and distance difference between nodes, the longer the communication connection between nodes is maintained, and the higher the link connectivity. The larger; Based on the link connectivity between nodes Provides a measure of the global link connectivity of the vehicle group. The details are as follows: Self-driving car swarm The trajectory separation is expressed as: in, (Global Trajectory Similarity) represents the global trajectory similarity of an autonomous vehicle ensemble. (Global Link Connectivity) represents the global link connectivity of an autonomous vehicle swarm. and These are the normalization coefficients for scaling the corresponding dimensions. ; Step 3.3 Vehicle Role To improve the stability of the autonomous vehicle swarm, a system with a master leader node and auxiliary coordinating nodes is adopted to achieve coordination and intelligent interaction among multiple vehicle systems. The composition of the vehicle swarm and its members is as follows: (1) Autonomous vehicle swarm G: The autonomous vehicle swarm is represented by the following quintuple: in, This represents the set of autonomous vehicle nodes within a vehicle group. This represents the set of directed edges that maintain communication connections between nodes within a vehicle group, where the directed edges... Represents a node Through nodes Join the vehicle group and maintain communication connectivity; This represents the set of node roles for autonomous vehicles, including the lead node. Collaborative nodes Follow the node and edge nodes ; A function representing the mapping from nodes to roles within a vehicle group; express The set of target nodes for directed edges, namely the set of leading nodes and collaborating nodes; (2) Lead Node (LN): When an edge node... The number of neighborhood sets is greater than the threshold , The election of a leading node is conducted when The leading probability function value If it is higher than all edge nodes in its neighborhood set, then Called the leading node, it can be represented as: in, This represents the threshold that the number of neighbors a node must satisfy in order for an edge node to be able to elect a leader node; The lead node is responsible for communication, computing, and storage management of member information among vehicle group members, thereby maintaining the stability of the vehicle group; (3) Cooperative Node (CN): When an edge node There is a leading node in the neighborhood set. ,and Received leading node The message allowing joining the car group then... The coordinating nodes, referred to as the vehicle group, can be represented as: in, Represents a node Received leading node The message indicating agreement to join the car group; The cooperating nodes are responsible for the addition of follow nodes and preventing the re-clustering overhead caused by the departure of the leading node, thereby sharing the communication and computing load of the leading node and cooperating to maintain the stability of the vehicle group. (4) Follower node FN: When an autonomous vehicle There are cooperating nodes in the neighborhood set. ,and Received from the cooperating node The message allowing joining the car group then... A following node in a vehicle group can be represented as: (5) Edge Node (EN): When the edge node There are no leading or cooperating nodes in the neighborhood set of a given node, and the probability function value of its leading node is [value missing]. If it is not the largest neighborhood set, then This is called an edge node, and it represents the initial state of an autonomous vehicle, as follows: Step 3.4 Lead Node Election (1) Properties related to leader election 1) Trajectory Representation driverless vehicles The trajectory characterization degree characterizes The ability to represent the overall trajectory trend of its neighborhood set, namely: in, Represents a node The average trajectory similarity with all nodes in its neighborhood set, i.e.: Property 1: ; Property 1 represents driverless vehicles. The more similar the trajectory is to the node in the neighborhood set, the better. The greater the trajectory representation; 2) Motion similarity driverless vehicles Motion similarity characterizes The degree of similarity between the movement patterns of its neighboring nodes, i.e.: in, and These are nodes and nodes The difference between velocity and acceleration; From formula (32), we can see that, That is, when the velocity and acceleration between nodes are the same, the more similar the trends of motion changes of the nodes, the better. The larger; 3) Link connectivity driverless vehicles The link connectivity characterizes The persistence of the communication link between it and its neighboring nodes can be obtained according to formula (16): From formula (33), it can be seen that the link connectivity is negatively correlated with the speed difference and distance between the two nodes. Therefore The smaller the speed and distance differences between the nodes and the neighboring set nodes, the longer the communication connection between the nodes can be maintained. Link connectivity The larger; (2) Leader election algorithm driverless vehicles The probability of being elected as the leader node is formally represented using a normalized probability function, which comprehensively considers the autonomous vehicle. The trajectory representation degree, motion similarity, and link connectivity of the relative autonomous driving neighbor list are as follows: in, , and These are the normalization coefficients; In summary, in the list of autonomous driving neighbors, autonomous vehicles The more similar the trajectory and movement pattern to the list of autonomous driving neighbors, the greater the link connectivity. Leading probability function The larger the value, The higher the probability of being selected as the leading node; Step 3.5 Vehicle Group Formation Algorithm After the successful construction of the node neighborhood set and the completion of the election of the leading node, all nodes have their own relatively stable neighborhood set information.

5. The method for forming an autonomous vehicle swarm based on trajectory prediction as described in claim 4, characterized in that, Step 3.4 Lead Node Election: The leader node election process based on the leader probability function is shown in Algorithm 4, and the specific steps are as follows: (1) All nodes are initialized as edge nodes and broadcast HELLO data packets to their surrounding neighbors; (2) After each node has the motion information of its neighboring nodes, it calculates its own leading function value according to formula (34). And broadcast the LEADER_PROB data packet; (3) If the number of neighboring nodes in the neighborhood set of node P reaches the threshold Node P sorts the leading probability function values ​​of its neighboring nodes in the neighborhood set. If node P has the largest leading probability function value, then node P is selected as the leading node; otherwise, it retains its status as an edge node and broadcasts a request to join the vehicle group message JOIN_REQ.

Images