Lane changing and overtaking assistance method for networked vehicles in mixed sparse scenarios

Abstract

The application provides a lane changing and overtaking assisting method for networked vehicles in a mixed sparse scene, comprising the following steps: a networked vehicle real-time senses state information of traditional vehicles on a road and stores the state information; the networked vehicle traverses state information of surrounding traditional vehicles, identifies missing time stamps or corresponding state information, and repairs the missing time stamps or corresponding state information; a vehicle state prediction model is established and the vehicle state prediction model is used to predict future states of the networked vehicle and surrounding vehicles; a lane changing and overtaking model is established and the lane changing and overtaking model is used to simulate and test a lane changing and overtaking process in advance; results output by the lane changing and overtaking model are used to evaluate feasibility of a vehicle in performing a lane changing and overtaking action at a future time, and speed guidance corresponding to lane changing and overtaking is provided to the vehicle with higher feasibility. The lane changing and overtaking assisting method for networked vehicles can assist the networked vehicle to complete lane changing and overtaking operation, and solves the problem of how the networked vehicle safely and efficiently performs lane changing and overtaking in a mixed sparse scene.

Description

Technical Field

[0001] This invention belongs to the field of vehicle wireless communication technology, and in particular relates to a method for assisting networked vehicles in lane changing and overtaking in a hybrid sparse scenario. Background Technology

[0002] With the continuous development of Intelligent Transportation Systems (ITS), innovations in technologies such as traffic signal control and advanced driver assistance systems have greatly improved the efficiency and safety of road traffic. The development of ITS has also driven the emergence of autonomous vehicles, providing prospects for solving modern traffic problems and realizing intelligent mobility. As a key technology for enhancing vehicle intelligence, Vehicle-to-Everything (V2X) communication technology primarily enables vehicles equipped with On-Board Units (OBUs) to establish connections with their entire environment. V2X communication technology has become a key driver for improving road traffic efficiency. However, due to multiple factors such as technical standards, infrastructure requirements, and costs, the adoption of V2X vehicles has been relatively slow. Therefore, in the foreseeable future, roads will be shared by two types of vehicles: connected vehicles equipped with OBUs and traditional vehicles without OBUs. This hybrid approach presents challenges because it hinders direct communication between these two types of vehicles. For example, in a reversing overtaking scenario, traditional vehicles cannot communicate with vehicles coming from the opposite direction. Therefore, changing lanes in the presence of blind spots becomes dangerous, potentially leading to collisions. Furthermore, connected vehicles need to comply with traffic regulations and driving habits. When a vehicle intends to change lanes, it needs to send a lane change request to surrounding vehicles (e.g., activating its turn signal). The actual lane change is only carried out after a certain period of time has elapsed. This advance warning to surrounding vehicles can reduce the risk of collision to some extent.

[0003] Lane changing and overtaking are common driving behaviors on roads. Appropriate lane changes based on current traffic conditions can significantly improve driving efficiency and alleviate traffic congestion. Conversely, impatient overtaking, ignoring nearby vehicles, and frequent lane changes are considered unreasonable lane-changing behaviors, leading to traffic chaos and negatively impacting overall traffic flow stability. Therefore, assisting connected vehicles in performing appropriate lane-changing behaviors is crucial for maintaining smooth and safe road traffic. Connected vehicles typically delay executing lane changes after sending lane-change requests to surrounding vehicles. This time allows for the effective utilization of the time to assess the feasibility of subsequent overtaking maneuvers. This provides decision-making references for connected vehicles, improving lane-changing and overtaking efficiency. Existing technologies in this field focus on globally automated driving scenarios, assuming all vehicles can communicate with each other and can change lanes at any time without first signaling to nearby vehicles. However, in mixed scenarios, this lane-changing approach poses a significant collision risk to surrounding vehicles. When a connected vehicle senses the state of surrounding vehicles within its perception range, state data frames may be lost due to factors such as vehicle obstruction. Ensuring the integrity of perception data can reduce erroneous predictions of future states. The future state of a vehicle may be potentially influenced by future driving behaviors. Different driving behaviors are typically accompanied by different position, speed, and acceleration patterns, which provide essential information for more accurate predictions of a vehicle's future state. Furthermore, traffic flow in each lane ahead of a connected vehicle affects its ability to change lanes effectively. For example, if a vehicle ahead is obstructing traffic, a connected vehicle may consider accelerating to change lanes. However, if the vehicle ahead is still moving slowly or the lane is congested after the lane change, changing lanes may not be efficient, and frequent lane changes may also pose risks and slow down traffic flow.

[0004] Existing known technologies mainly categorize these into vehicle future state prediction and lane-changing / overtaking. Among known technologies for predicting vehicle future states, the focus is primarily on predicting the vehicle's future position (i.e., trajectory). Hou et al., in "Interactive Trajectory Prediction of Surrounding Road Users for Autonomous Driving Using Structural-LSTM Network," introduced a hierarchical multi-sequence learning network that automatically captures the complex dependencies between multiple interacting vehicles using a structured long short-term memory (LSTM) network to accurately predict the future trajectories of surrounding vehicles. Messaoud et al., in "Attention Based Vehicle Trajectory Prediction," employed a vehicle interaction model-based approach, drawing on human reasoning and introducing an attention mechanism to emphasize the critical influence of surrounding vehicles on the future state, thus solving the trajectory prediction problem for autonomous driving on highways. Hou et al., in "Integrated Graphical Representation of Highway Scenarios to Improve Trajectory Prediction of Surrounding Vehicles," proposed a unified graphical representation method that comprehensively considers various factors such as different numbers and types of vehicles, road structure, and traffic rules to improve the accuracy of trajectory prediction. Specific methods include assessing road and vehicle passability, defining collision risk levels, rasterizing the scene and generating feature maps, and finally performing trajectory prediction using an LSTM network. Besides predicting the future position of vehicles, a large amount of research focuses on predicting lane-changing behavior. Wang et al., in "An Intelligent Lane-Changing Behavior Prediction and Decision-Making Strategy for an Autonomous Vehicle," proposed a model for predicting vehicle lane-changing behavior. This model includes a base model based on LSTM and an adaptive model with an adaptive decision threshold embedded in the base model. Bayesian inference is used to rapidly update the threshold.

[0005] In another well-known technique for lane changing and overtaking, the lane changing and overtaking problem is an extremely complex decision-making scenario that requires effective management of the interactions and nonlinear relationships between multiple vehicles. Deep reinforcement learning (DRL) has the ability to learn, adapt, and continuously enhance decision-making strategies to adapt to different traffic and road conditions. Therefore, DRL is often used to solve the lane changing and overtaking problem. Zhang et al., in "Multi-Agent DRL-Based Lane Change With Right-of-Way Collaboration Awareness," proposed a two-layer lane change planning strategy. This strategy first uses a multi-agent DRL model to redefine the lane changing problem so that vehicles can cooperate better. Second, it incorporates the driving intentions of surrounding vehicles into the model to implicitly negotiate right-of-way, consider traffic impacts, and guide the coordinated movement of the multi-vehicle system. Chen et al., in "Automatic Overtaking on Two-way Roads with Vehicle Interactions Based on Proximal Policy Optimization," used a proximal policy optimization method to formulate overtaking strategies for autonomous vehicles. By introducing a hierarchical control framework for road vehicle management, this framework monitors vehicle behavior at a high level and manages vehicle movement at a low level. The optimization algorithm is used to generate advanced decision-making strategies, and a self-attention mechanism is introduced to improve the algorithm's performance.

[0006] Analysis reveals several key limitations in the relevant prior art technologies: First, most of these technologies focus on improving neural network models to enhance the accuracy of future vehicle state predictions, often neglecting the significant impact of future driving behavior on state prediction. Second, existing technologies focus on global autonomous driving, assuming vehicles can obtain accurate and complete information about the states of surrounding vehicles and that vehicles can communicate seamlessly. However, this assumption is unrealistic in many situations. Third, in these prior art technologies, lane changes occur in real time without requiring lane change requests to surrounding vehicles. However, in mixed scenarios, this approach can pose significant safety risks. Finally, most prior art technologies ignore traffic flow conditions in specific areas ahead of the vehicle, potentially leading to unnecessary lane changes.

[0007] Therefore, a method for lane changing and overtaking by connected vehicles has been proposed to address the challenges brought about by hybrid driving scenarios, which has become an urgent problem to be solved. Summary of the Invention

[0008] In view of this, the present invention discloses a method for assisting connected vehicles in lane changing and overtaking in a mixed sparse scenario, so as to solve the problem of how connected vehicles can safely and efficiently perform lane changing and overtaking actions in a sparse mixed scenario.

[0009] This invention provides a method for assisting connected vehicles in lane changing and overtaking in a hybrid sparse scenario, comprising the following steps:

[0010] S1: The connected vehicle equipped with OBUs can perceive and store the status information of traditional vehicles on the road that are not equipped with OBUs in real time. The status information includes x, y, v, a, (x, y) representing the vehicle's position on the road, v representing the vehicle's speed, and a representing the vehicle's acceleration.

[0011] S2: The connected vehicle traverses the status information of surrounding traditional vehicles, identifies lost timestamps or corresponding status information, and repairs the missing status information by combining the vehicle kinematics model.

[0012] S3: Establish a vehicle state prediction model and use the vehicle state prediction model to predict the future state of the connected vehicle and its surrounding vehicles. The input of the vehicle state prediction model includes the vehicle's historical state information, predicted driving behavior information, and the interaction relationship between vehicles. The output of the vehicle state prediction model is the predicted state information of the connected vehicle and its surrounding vehicles when the connected vehicle actually performs a lane change and overtaking maneuver in the future.

[0013] S4: Establish a lane change and overtaking model and use the lane change and overtaking model to simulate and test the lane change and overtaking process in advance. The lane change and overtaking model adopts an improved parameterized deep Q-learning network to handle the discrete actions of lateral lane change and the continuous actions of longitudinal speed control involved in the lane change and overtaking process.

[0014] S5: Use the results output by the lane change and overtaking model to evaluate the feasibility of the vehicle performing a lane change and overtaking maneuver at a future time, and provide corresponding speed guidance for lane change and overtaking for vehicles with higher feasibility.

[0015] Preferably, S2 specifically includes the following steps:

[0016] S21: The connected vehicle scans the status information of surrounding traditional vehicles and identifies the lost timestamps or corresponding status information.

[0017] S22: If the lost timestamp is determined, select (x, y, v, a) values ​​from the times before and after the lost timestamp, and then combine the values ​​of these two times with the vehicle kinematics principle to calculate the missing state data;

[0018] S23: Repair the vehicle's status information using the calculated missing status data.

[0019] Further optimization, in S3, the method for establishing the interaction relationship between vehicles is as follows:

[0020] Based on the location (x) of the connected vehicle t ,y t Determine the vehicle ID for each region;

[0021] Obtain historical status information of the corresponding vehicle from the data collected and stored from connected vehicles;

[0022] By comparing the location of connected vehicles in historical status information with the status data of each vehicle in each region, the difference between the lateral and longitudinal coordinates of each vehicle at each time step is calculated, thereby establishing the interaction relationship between vehicles and determining the value of each attribute in relPos.

[0023] Further optimization, in S3, the method for predicting the future driving behavior of vehicles is as follows: the future driving behavior of vehicles is predicted by utilizing the interaction relationship between vehicles. An LSTM-based driving behavior prediction model is proposed. Each independent LSTM model is associated with the networked vehicle and its surrounding related vehicles. The hidden state and cell state of the LSTM are connected through radial connections to understand the high-level dependencies between interacting vehicles, realize the sharing of hidden states, and thus predict the future driving behavior of vehicles.

[0024] Further optimization, in S3, the process of predicting the future state of the vehicle is as follows:

[0025] A three-dimensional spatial grid G ​​is constructed based on the location of the connected vehicle at time t, consisting of the hidden states of the connected vehicle and its surrounding vehicles.

[0026] By embedding the state vector of each vehicle through a fully connected layer, an embedding vector covering the adjacent region is generated. An attention mechanism is introduced to assign weights to different locations and related features.

[0027] The feature vectors generated by the attention mechanism are sent to another fully connected layer for further processing and feature integration;

[0028] The results of cascading LSTM and multi-head attention mechanisms generate more comprehensive representations;

[0029] The fully connected layer performs decoding and outputs the vehicle state prediction result.

[0030] Further optimization, in S4, the establishment of the lane-changing and overtaking model includes the following steps:

[0031] S41: The vehicle's actions during the entire lane-changing and overtaking process are abstracted into discrete lane-changing actions and continuous speed control;

[0032] S42: Design a lane-changing and overtaking model, which adopts an actor-critic structure. In this structure, the actor network uses the PPO algorithm, and the critic network uses Dueling DQN.

[0033] Further optimization in S5 involves four basic factors: the feasibility of guiding connected vehicles to the corresponding speed, the risk status, the number of lane changes required to complete overtaking, and the accuracy of predicting the status of surrounding vehicles.

[0034] Further optimization, in S5, the specific definition of the Feasibility Analysis Index (FAI) is as follows:

[0035]

[0036] Where, k v ,k t ,k l and k r These represent the weights of FAI in terms of speed maneuverability, risk status, number of lane changes, and prediction error, respectively. The subscripts t and i represent the number of surrounding vehicles in the study area at each time step, respectively, and t represents the maximum time step. v is used as the unit of measurement. t+1 -v t The system measures the change in vehicle speed between two moments before and after an action. The max function ensures that the calculation results of the four factors fall within a similar range. RMSE represents the root mean square error of the state of the connected vehicle and its surrounding vehicles at the prediction time t+t'.

[0037] This invention provides a method for assisting connected vehicles in lane changing and overtaking in hybrid sparse scenarios. Given the reality that the penetration rate of connected vehicles on roads cannot quickly reach 100%, this invention is the first to consider how to assist connected vehicles in lane changing and overtaking in real-world hybrid scenarios. First, the connected vehicle uses multiple onboard sensors to jointly perceive the state of surrounding vehicles and record state data. For situations where state data is lost due to occlusion, this data is repaired and filled using a vehicle kinematics model. After the connected vehicle issues a lane change request, it predicts the state of its own vehicle and surrounding vehicles. Next, the predicted state is mapped to the initial state of the lane changing and overtaking model, and the feasibility of the vehicle performing the lane changing and overtaking maneuver is evaluated by comprehensively considering factors such as safety, efficiency, and traffic flow ahead. Finally, a feasibility evaluation index is designed, comprehensively considering factors such as the accuracy of state prediction and the rate of speed change during the overtaking process, providing a decision-making reference for executing the lane changing and overtaking maneuver.

[0038] Therefore, compared with existing related technologies, this solution addresses the challenge of non-real-time lane changes by vehicles in mixed scenarios by comprehensively considering multiple factors, such as vehicle driving behavior and lane traffic dynamics.

[0039] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit the disclosure of the present invention. Attached Figure Description

[0040] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of the method for assisting networked vehicles in lane changing and overtaking in a hybrid sparse scenario provided by the present invention;

[0043] Figure 2 This is an overall framework diagram of the method for assisting networked vehicles in lane changing and overtaking in a hybrid sparse scenario provided by the present invention.

[0044] Figure 3 This is an architecture diagram for predicting the future state of vehicles;

[0045] Figure 4 This is a diagram illustrating the architecture of vehicles changing lanes and overtaking. Detailed Implementation

[0046] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of systems consistent with some aspects of the invention as detailed in the appended claims.

[0047] like Figure 1 As shown, this invention provides a method for assisting connected vehicles in lane changing and overtaking in a hybrid sparse scenario, aiming to safely and efficiently assist connected vehicles in lane changing and overtaking operations, including the following steps:

[0048] S1: Connected vehicles equipped with OBUs (on-board units) use their own onboard sensor devices to perceive and store the status information of traditional vehicles on the road that are not equipped with OBUs in real time. The status information includes x, y, v, a, where (x, y) represents the vehicle's position on the road, v represents the vehicle's speed, and a represents the vehicle's acceleration.

[0049] S2: The connected vehicle traverses the status information of surrounding traditional vehicles, identifies lost timestamps or corresponding status information, and repairs the missing status information by combining the vehicle kinematics model.

[0050] Because real-world road conditions may cause intermittent blockages in the sensors used for collecting status information, affecting the sensors' continuous perception of the status of surrounding vehicles, there may be gaps and omissions in some status information. Therefore, it is necessary to fill in the gaps left by missing data.

[0051] S3: Establish a vehicle state prediction model and use the vehicle state prediction model to predict the future state of the connected vehicle and its surrounding vehicles. The input of the vehicle state prediction model includes the vehicle's historical state information, predicted driving behavior information, and the interaction relationship between vehicles. The output of the vehicle state prediction model is the predicted state information of the connected vehicle and its surrounding vehicles when the connected vehicle actually performs a lane change and overtaking maneuver in the future.

[0052] Specifically, the interaction relationships between vehicles are analyzed first. The vehicle ID for each region is determined by the location of the connected vehicles. Then, historical state information of the corresponding vehicles is obtained from the data collected and stored by the connected vehicles. By comparing the location of the connected vehicles in the historical state information with the state data of each vehicle in each region, the differences in the lateral and longitudinal coordinates of each vehicle at each time step are calculated, thereby establishing the interaction relationships between vehicles. Next, this relationship is used to predict the future driving behavior of the vehicles. The input to the state prediction model includes the vehicle's historical information, the predicted driving behavior, and the interaction information between vehicles. The output of the model is the predicted state of the connected vehicle, the target vehicle, and surrounding vehicles when the connected vehicle actually performs a lane change and overtaking maneuver in the future.

[0053] S4: Establish a lane change and overtaking model and use the lane change and overtaking model to simulate and test the lane change and overtaking process in advance. The lane change and overtaking model adopts an improved parameterized deep Q-learning network to handle the discrete actions of lateral lane change and the continuous actions of longitudinal speed control involved in the lane change and overtaking process.

[0054] Specifically, the model abstracts the vehicle's actions throughout the lane-changing and overtaking process into a discrete lane-changing action and a continuous speed control, and then parameterizes them. The model design employs an actor-critic structure. The connected vehicle is considered an intelligent agent, with the algorithm controlling the vehicle's lane changes and speed adjustments. To better simulate human drivers, IDM and MOBIL models are used to control the motion of traditional vehicles.

[0055] S5: Use the results output by the lane change and overtaking model to evaluate the feasibility of a vehicle performing a lane change and overtaking maneuver at a future time, and provide corresponding lane change and overtaking speed guidance for vehicles with higher feasibility.

[0056] This method for assisting connected vehicles in lane changing and overtaking in a hybrid sparse scenario aims to predict the future state of vehicles and perform feasibility analysis for lane changing and overtaking. The method first describes the process by which connected vehicles perceive and process the states of surrounding vehicles on the road. Then, this perceived data is used as the basis for predicting the future states of these vehicles when the connected vehicle actually performs a lane changing and overtaking maneuver. After predicting the vehicle states, the predicted states are mapped to the initial states of the lane changing and overtaking model. Finally, the feasibility of lane changing and overtaking is evaluated using the predicted vehicle states and the lane changing and overtaking model.

[0057] The overall framework of the present invention is as follows Figure 2 As shown, this invention addresses the decision-making process for connected vehicles performing lane-changing and overtaking maneuvers. To illustrate this process in detail, four key vehicles are focused on: the CV (the vehicle performing the lane-changing and overtaking maneuver), the target vehicle TV (the intended overtaking vehicle), and two surrounding vehicles SVs. In the initial phase, the CV uses its sensors to jointly perceive the states of TV and SVs. The collected data is then processed to construct a historical state set for the vehicles. In the intermediate phase, the historical state set is used to predict the states of these four vehicles when the CV actually performs an overtaking maneuver in the future. For example, at time t, the CV attempts to gain higher speeds by changing lanes and overtaking. In this scenario, involving both connected and conventional vehicles, the CV first issues a lane-changing request, alerting SVs and TV. The actual lane-changing action occurs at a later time t', at which point the states of the four vehicles are predicted. In the final phase, the predicted states are mapped to the initial states of the lane-changing and overtaking model. Then, the trained lane-changing and overtaking model deployed in the CV is used to simulate the lane-changing and overtaking maneuver in advance. Furthermore, during the decision-making process, the CV considers traffic flow within a certain range ahead of the vehicle as an influencing factor. When a connected vehicle needs to actually perform a lane change and overtake maneuver, the CV will consider whether to proceed with the maneuver based on a feasibility index. A high feasibility assessment index (FAI) indicates a greater likelihood of successfully completing the overtake maneuver and will provide speed guidance to the CV to help complete the lane change and overtake maneuver.

[0058] S2 specifically includes the following steps:

[0059] S21: The connected vehicle scans the status information of surrounding traditional vehicles and identifies the lost timestamps or corresponding status information.

[0060] S22: If the lost timestamp is determined, select (x, y, v, a) values ​​from the times before and after the lost timestamp, and then combine the values ​​of these two times with the vehicle kinematics principle to calculate the missing state data;

[0061] S23: Repair the vehicle's status information using the calculated missing status data.

[0062] Given three consecutive time points, the state at the time point where intermediate state data is missing is calculated using the states of the two consecutive time points. The calculation method is as follows:

[0063]

[0064] In the formula, (preX, preY), preV and preA represent the position, velocity and acceleration at the previous moment, and (nextX, nextY), nextV and nextA represent the position, velocity and acceleration at the next moment.

[0065] Furthermore, to improve computational efficiency, connected vehicles limit the maximum length of status information for each vehicle in their vicinity to 60 bytes, meaning they only store the status data for each vehicle for the most recent 6 seconds. This limitation primarily takes into account the limited computing and storage capabilities of OBUs.

[0066] S3 includes the following steps: establishing inter-vehicle interaction relationships, predicting future vehicle driving behavior, and predicting future vehicle states.

[0067] The method for establishing the interaction relationships between vehicles is as follows:

[0068] The interaction between vehicles depends on six attributes, denoted as relPos i,t ={(LF i,t ,LB i,t ,MF i,t MB i,t ,RF i,t ,RB i,t )|i,t∈(0,1...)}, where relPos i,t This represents the relative position coordinates of surrounding vehicles within six adjacent areas of connected vehicle i at time t. First, based on the position (x) of the connected vehicle... t ,y t The system determines the vehicle ID for each region, then retrieves the historical status information of the corresponding vehicle from the data collected and stored by the networked vehicles. Next, it compares the location of the networked vehicle in the historical status information with the status data of each vehicle in each region, calculates the difference between the lateral and longitudinal coordinates of each vehicle at each time step, and then establishes the interaction relationship between the vehicles, thus determining the value of each attribute in relPos.

[0069] The method for predicting future driving behavior of vehicles is as follows: An LSTM-based driving behavior prediction model is proposed, where each independent LSTM model is associated with a connected vehicle and its corresponding surrounding vehicles. The hidden states and unit states of the LSTMs are connected radially to understand the high-level dependencies between interacting vehicles, enabling the sharing of hidden states and thus predicting future driving behavior. The vehicle's state over a past period may reflect future maneuverability to some extent; for example, multiple movements at a certain heading angle suggest a possible lane-changing maneuver. Furthermore, vehicle driving behavior not only affects its own state but may also influence the state of surrounding vehicles to some extent. For example, if CV1 is driving normally on the road while CV2 suddenly changes lanes and enters the lane in front of CV1, CV1 may brake to avoid a collision. Therefore, predicting the vehicle's potential driving behavior is crucial before predicting its future state. To predict future driving behavior, an LSTM-based driving behavior prediction model is proposed. Each individual LSTM model is associated with a connected vehicle and its corresponding surrounding vehicles. These LSTM models are designed to capture the spatiotemporal features of each vehicle's historical state. Then, these adjacent LSTM models are connected, allowing for the sharing of hidden states for joint prediction of future behavior. By radially connecting the hidden states and unit states of the LSTM, high-level dependencies between interacting vehicles can be understood. Although the LSTMs associated with vehicles in the six defined regions may differ in input length, they all share the same input and output types. This interconnected LSTM framework improves the model's ability to capture the collaborative dynamics between vehicles, thereby improving the accuracy of driving behavior prediction. For each state, each vehicle within the networked vehicle interaction range and its state S over the past n time steps are considered. i,t-j ={(x i ,y i ,v i ,a i ,θ i ) | i∈(0,1,2…), j∈(0,1,2…)} and relative state information relPos i The data is input into the first layer of the model. The second layer consists of an LSTM. The historical update state h' of vehicle i is calculated over multiple time steps. i The hidden state vector is fed into the LSTM model. During each time step, the LSTM generates a hidden state vector, as shown below:

[0070] h″ i,t =LSTM(h′) i,t ,h″ i,t-1 W lstm (2)

[0071] Among them, W lstmThis represents the learnable parameters of the LSTM model, used to update the hidden state vector. D l h″ represents the dimension of the hidden state vector. i,t-1 This represents the previous hidden state vector. The final prediction of future vehicle behavior is achieved through a combination of LSTM layers and fully connected (FC) layers. The output is y. i As shown below:

[0072] y i =Linear(h″) i,t W fc (3)

[0073] W fc These are the learnable parameters of the FC layer.

[0074] The process of predicting the future state of a vehicle is as follows:

[0075] The inputs to the vehicle state prediction model include historical vehicle information, predicted driving behavior, and interactions between vehicles, which are expressed as follows:

[0076]

[0077] Where i represents the corresponding vehicle number, and N represents the maximum number of predicted vehicles. The output of the vehicle state prediction model is the predicted state of the connected vehicle and surrounding vehicles when the vehicle actually performs a lane change and overtaking maneuver in the future. Output state Y i,t+t′ =(x i ,y i ,v i ,a i ), where i∈(1,N). The specific output corresponds to Figure 3 The table in the bottom right corner.

[0078] Specifically, firstly, a three-dimensional spatial grid G ​​is constructed based on the location of the connected vehicle at time t. This grid consists of the hidden states of the connected vehicle and surrounding vehicles at time t, represented as:

[0079]

[0080] function τ mn The (x,y) parameter determines whether the position (x,y) is within a grid cell (m,n). If it is, a value of 1 is assigned; otherwise, a value of 0 is assigned. TThe selected area represents all surrounding vehicles, and the grid consists of m rows, n columns, and the hidden state of each vehicle. The grid's extent is centered on the location of the networked vehicle, and its size is represented as (M, N). The grid primarily covers a horizontal distance equivalent to three lanes (meters) and a vertical distance spanning five vehicle lengths (meters). Then, vehicle i is embedded into each state vector through an FC layer. To generate an embedding vector e that covers information about neighboring regions. t,i Over the past n time steps, the LSTM encoder receives the embedding vector of each vehicle i as input. The LSTM encoder then stores the previous hidden states... and the current embedding vector As input, to compute the current hidden state vector As shown in the following formula:

[0081]

[0082] Among them, W enc This represents the corresponding encoder weights shared by each LSTM encoder. Finally, the future state of the vehicle is predicted based on the above two steps. To effectively capture the spatiotemporal relationships between vehicles, this invention introduces an LSTM-based attention mechanism. This mechanism is mainly based on the spatiotemporal encoding between connected vehicles and surrounding vehicles. The network structure of this method includes the following key steps: First, the initial input data X... i The data is fed into an LSTM layer for encoding. Then, the output of the LSTM layer is further processed and mapped to the encoded features through an FC layer, transforming the state information into a higher-dimensional representation. To capture important features and relationships in vehicle states, and considering that vehicles on the road are not limited to single interconnected groups, this method introduces a multi-head attention mechanism. This mechanism can focus on multiple subgroups of vehicles. These subgroups represent different aspects of spatiotemporal relationships that may occur between multiple nearby vehicles, and summarize different nodes with different importance scores under each spatiotemporal relationship. Different sets of Q, K, and V matrices are generated in parallel across nh attention heads, denoted as Q... l ,K l and V l By mapping the hidden state of vehicles to queries Map grid to keywords Sum Among them W η,l W θ,l and W κ,lThis is the weight matrix learned in each attention head. η is used to map the hidden state to Q, θ is used to map the grid G ​​to the set K, and κ is used to map the corresponding values ​​of surrounding vehicles to a set represented as V. This method employs dot product attention, primarily calculating the weights of the attention mechanism based on the interaction between a pair of vehicles. The specific formula for calculating the attention weights is shown below:

[0083]

[0084] Among them, Q l and K l These are the query vector and the key vector, respectively, where d is the scaling factor for the projection space dimension. Using Q... l and The dot product is used to calculate the similarity between the query and the key vector. Furthermore, V... l The system uses feature representations to represent information, such as the historical states of surrounding vehicles. Then, attention weights are calculated using a softmax function. Finally, the attention mechanism computes the encoded vectors of surrounding vehicles and performs a weighted sum, as shown in the formula below:

[0085]

[0086] in, The encoded vector representing the surrounding vehicle j. This represents the corresponding attention weight for the surrounding vehicle j.

[0087] In S4, such as Figure 4 As shown, vehicle lane-changing and overtaking involve two actions: lateral lane-changing and longitudinal speed control. This invention proposes the PPODDQN algorithm for mixed actions of both discrete and continuous types. This algorithm can effectively predict the feasibility of lane-changing and overtaking, reduce unnecessary lane-changing behaviors, and thus improve road traffic efficiency. It integrates an actor-critic structure, where the actor network is trained stably using PPO, and the critic network uses Dueling DQN (DDQN) to estimate state values, while parameterizing the action space. The core idea of ​​this method is to train the actor network using PPO, evaluate the state values ​​using DDQN, and use an n-step update method to address the delayed reward problem. This method treats the networked vehicle as an intelligent agent, and its algorithm controls the vehicle's lane changes and corresponding speed control. To simulate the behavior of a human driver, an Intelligent Driver Model (IDM) and a Minimize Overall Braking Caused by Lane Changes (MOBIL) model are used to control the movement of the traditional vehicle.

[0088] Specifically, the vehicle's state space mainly consists of the connected vehicle's state information and interaction information with surrounding vehicles. To describe the state of surrounding vehicles in detail, the surrounding environment is divided into six different sub-regions: left front, left rear, front directly, rear directly, right front, and right rear. The connected vehicle's state information includes the lane ID, longitudinal distance from the road start point, speed, and acceleration. The interaction information with surrounding vehicles includes the longitudinal distance difference with connected vehicles and the speed information of the corresponding vehicles. These parameters combine to form the state space. If there are no vehicles in a specific region, the speed and distance difference values ​​for that region are set to -1. It is important to note that the distance to vehicles outside the perception range is set to 150 meters. The actions performed by the connected vehicle are divided into discrete lateral lane changing actions and continuous longitudinal speed control. The entire action space is represented as follows:

[0089]

[0090] Where 0 represents lane keeping, -1 represents left lane changing, and 1 represents right lane changing. The algorithm's termination conditions are divided into two categories: reaching the specified time step and completing the overtaking maneuver. The termination conditions are as follows:

[0091]

[0092] Where, Δd i,tar Δv i,tar Let represent the longitudinal distance difference between the connected vehicle and the target vehicle to be overtaken, and the speed difference between the two vehicles, respectively. We designed a reward function that comprehensively considers traffic flow ahead, comfort, overtaking efficiency, safety risks, and execution costs. The reward function is expressed as follows:

[0093] r=ω1r frontTra +ω2r comf +ω3r effic +ω4r safe +ω5r negOver (11)

[0094] Among them, ω1, ω2, ω3, ω4, and ω5 are used to set the weight coefficients of the above five factors, respectively. Their specific descriptions are as follows:

[0095]

[0096]

[0097]

[0098]

[0099]

[0100] Where Q represents the number of vehicles passing through the lane, L represents the length of the lane within the study area, and t represents the time interval for calculating the traffic flow of each lane. min and a max These represent the vehicle's minimum and maximum accelerations, respectively. 'a' is defined based on the vehicle's actual driving performance characteristics. max =5m / s 2 ,a min = -5m / s 2 n is the number of lane changes, v target denoted by s, v represents the speed required for the vehicle to successfully complete an overtaking maneuver. α is an adjustable coefficient used to reduce the safety reward received by the agent that fails to complete the overtaking maneuver. The core idea of ​​the PPODDQN proposed in this invention is to parameterize the action space, associate each discrete action with a corresponding continuous action, and determine the optimal discrete and continuous actions hierarchically. The algorithm consists of two steps. First, the current input state is processed by the actor network (PPO) to generate the optimal continuous action. Second, the continuous actions generated in the first step are combined with the current state s by the critic network (DDQN) to calculate the dominance level of a series of discrete actions and their corresponding continuous actions. Finally, this process determines the optimal discrete action.

[0101] In S5, the feasibility of a vehicle performing a lane-changing and overtaking maneuver at a future point in time is assessed by comprehensively considering four basic factors: the speed feasibility of guiding the connected vehicle, the risk status, the number of lane changes required to complete the overtaking maneuver, and the error in predicting the vehicle's status. The specific feasibility analysis index is defined as follows:

[0102]

[0103] Where, k v ,k t ,k l and k r These represent the weights of vehicle maneuverability at speed, risk status, number of lane changes, and prediction error, respectively. The subscripts t and i represent the time step and the number of surrounding vehicles within the study area, respectively, and T represents the maximum time step. Using v... t+1 -v tThe FAI measures the change in vehicle speed between two moments before and after the maneuver. A larger change indicates decreased vehicle maneuverability and increased passenger discomfort. The max function ensures that the calculated results for the four factors fall within a similar range, allowing the FAI to fairly weigh these factors and avoid instability caused by any single factor being too large. RMSE represents the error in predicting the state of the connected vehicle and surrounding vehicles at time t+t′. The calculated FAI is provided as feedback to the connected vehicle to help it determine whether to perform a lane-change overtaking maneuver. If the connected vehicle chooses to perform the maneuver, speed guidance is provided to optimize the overtaking maneuver. A lower FAI suggests that staying in the current lane may be more prudent, as lane-change overtaking may present greater challenges or potential risks.

[0104] This invention proposes a lane-changing and overtaking assistance method for connected vehicles in mixed sparse scenarios. By providing connected vehicles with a non-binary lane-changing and overtaking feasibility assessment index and corresponding speed guidance, it assists connected vehicles in completing lane-changing and overtaking operations, solving the problem of how connected vehicles can safely and efficiently perform lane-changing and overtaking in mixed sparse scenarios. This further ensures road safety when connected vehicles and traditional vehicles coexist. Furthermore, this invention provides strong support for the integration of intelligent transportation systems and V2X communication technology.

[0105] This method begins with connected vehicles equipped with multiple sensors jointly sensing and recording the status of surrounding vehicles. To compensate for data loss due to occlusion, a vehicle kinematics model is used to repair and fill in the missing data. After a connected vehicle issues a lane change request, it predicts the status of itself and surrounding vehicles. Then, the predicted state is mapped to the initial state of the lane change / overtaking model. Taking into account factors such as safety, efficiency, and traffic flow ahead, a feasibility evaluation index is proposed to assess the feasibility of the lane change / overtaking maneuver, providing a decision-making reference for executing the maneuver.

[0106] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these changes and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for assisting networked vehicles in lane changing and overtaking in a hybrid sparse scenario, characterized in that, Includes the following steps: S1: Connected vehicles equipped with OBUs perceive and store the status information of traditional vehicles on the road that are not equipped with OBUs in real time, wherein the status information includes , , , , , Indicates the vehicle's position on the road. Indicates the speed of the vehicle. Indicates the vehicle's acceleration; S2: The connected vehicle traverses the status information of surrounding traditional vehicles, identifies lost timestamps or corresponding status information, and repairs the missing status information by combining the vehicle kinematics model. S3: Establish a vehicle state prediction model and use the vehicle state prediction model to predict the future state of the connected vehicle and its surrounding vehicles. The input of the vehicle state prediction model includes the vehicle's historical state information, predicted driving behavior information, and the interaction relationship between vehicles. The output of the vehicle state prediction model is the predicted state information of the connected vehicle and its surrounding vehicles when the connected vehicle actually performs a lane change and overtaking maneuver in the future. S4: Establish a lane change and overtaking model and use the lane change and overtaking model to simulate and test the lane change and overtaking process in advance. The lane change and overtaking model adopts an improved parameterized deep Q-learning network to handle the discrete actions of lateral lane change and the continuous actions of longitudinal speed control involved in the lane change and overtaking process. S5: Use the results output by the lane change and overtaking model to evaluate the feasibility of the vehicle performing a lane change and overtaking maneuver at a future time, and provide corresponding speed guidance for lane change and overtaking for vehicles with high feasibility. In S3, the method for establishing the interaction relationships between vehicles is as follows: Based on the location of connected vehicles Determine the vehicle ID for each region; Obtain historical status information of the corresponding vehicle from the data collected and stored from connected vehicles; By comparing the location of connected vehicles in historical status information with the status data of each vehicle in each region, the difference between the lateral and longitudinal coordinates of each vehicle in each time step is calculated, thereby establishing the interaction relationship between vehicles and determining the value of each attribute in relPos. In S3, the method for predicting the future driving behavior of vehicles is as follows: the interaction relationship between vehicles is used to predict the future driving behavior of vehicles. An LSTM-based driving behavior prediction model is proposed. Each independent LSTM model is associated with the networked vehicle and its surrounding related vehicles. The hidden state and cell state of the LSTM are connected through radial connection to understand the high-level dependency relationship between the interacting vehicles, realize the sharing of hidden state, and thus predict the future driving behavior of the vehicle. In S3, the process of predicting the future state of the vehicle is as follows: A three-dimensional spatial grid G ​​is constructed based on the location of the connected vehicle at time t, consisting of the hidden states of the connected vehicle and its surrounding vehicles. By embedding the state vector of each vehicle through a fully connected layer, an embedding vector covering the adjacent region is generated. An attention mechanism is introduced to assign weights to different locations and related features. The feature vectors generated by the attention mechanism are sent to another fully connected layer for further processing and feature integration; The results of cascading LSTM and multi-head attention mechanisms generate more comprehensive representations; The fully connected layer performs decoding and outputs the vehicle state prediction result. In S4, the establishment of the lane-changing and overtaking model includes the following steps: S41: The vehicle's actions during the entire lane-changing and overtaking process are abstracted into discrete lane-changing actions and continuous speed control; S42: Design a lane-changing and overtaking model, which adopts an actor-critic structure, in which the actor network uses the PPO algorithm and the critic network uses Dueling DQN; In S5, the feasibility analysis consists of four basic factors: the feasibility of guiding connected vehicles to the corresponding speed, the risk status, the number of lane changes required to complete overtaking, and the accuracy of predicting the status of surrounding vehicles. In S5, the Feasibility Analysis Index (FAI) is defined as follows: ； in, and , representing the weights of FAI in speed maneuverability, risk status, number of lane changes, and prediction error, respectively. The subscripts t and i represent the time step and the number of surrounding vehicles in the study area, respectively. To represent the maximum time step, use To measure the change in vehicle speed between two moments before and after an action, the max function is used to ensure that the calculated results of the four factors fall within a similar range. RMSE represents the speed difference between the connected vehicle and its surrounding vehicles at the predicted time. The root mean square error of the state.

2. The method for assisting networked vehicles in lane changing and overtaking in a hybrid sparse scenario according to claim 1, characterized in that: S2 specifically includes the following steps: S21: The connected vehicle scans the status information of surrounding traditional vehicles and identifies the lost timestamps or corresponding status information. S22: If the lost timestamp is determined, select (x, y, v, a) values ​​from the times before and after the lost timestamp, and then combine the values ​​of these two times with the vehicle kinematics principle to calculate the missing state data; S23: Repair the vehicle's status information using the calculated missing status data.

Images