Simulation model, method and system for lane changing behavior of human-driven vehicle

Abstract

The invention belongs to the technical field of automatic driving simulation, discloses a human-driven vehicle lane changing behavior simulation model, method and system, and aims to solve the problems that'pseudo-interaction 'interference exists in human-driven vehicle lane changing modeling, global'latest lane changing point' pressure conflicts with local vehicle dynamic response, and a traditional model is low in nonlinear heterogeneous behavior description precision. An LA-Transform lane changing behavior simulation model is designed, and the LA-Transform lane changing behavior simulation model is constructed through a collaborative architecture network of input embedding and feature preprocessing optimization, short-time dynamic feature extraction and integration of a long and short-term memory network, sparse optimization of a self-adaptive sparse attention mechanism and encoder-decoder global integration. The method is specially used for high-precision simulation of the lane changing behavior of people driving vehicles in the confluence area. Data used for model training are all from a CitySim public data set Freeway C scene, and experimental verification shows that the model can generate a track conforming to real driver lane changing characteristics.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving simulation technology, specifically relating to a simulation model, method, and system for lane-changing behavior of human-driven vehicles. Background Technology

[0002] In recent years, with the continuous growth of highway traffic flow dominated by human-driven vehicles, forced lane change scenarios require simultaneous control of longitudinal and lateral speeds to complete lane transfers. Furthermore, influenced by the spatial constraints of the "latest lane change point," human-driven behavior exhibits strong spatiotemporal dependence and uncertainty. Accurate modeling of human-driven behavior in this scenario is the core foundation for supporting high-fidelity simulation of highway traffic flow, human-driven behavior analysis, and traffic safety assessment. However, accurate modeling still faces significant challenges: information about vehicles in front and behind the target lane is often dynamically missing in the initial acceleration phase due to sparse traffic flow. Traditional models, such as the standard Transformer, rely on dense self-attention mechanisms, which still forcibly calculate interaction weights for missing data, creating "pseudo-interaction" noise. This increases the computational burden and dilutes the weights of key features such as the target lane's safety clearance, leading to biased judgments on the safety boundaries of human-driven lane changes. Simultaneously, forced lane change behavior is driven by both the global constraint of the "latest lane change point" and the local dynamics of surrounding vehicles. This contradiction between global constraints and local responses is prominent in various highway lane change scenarios, and traditional models struggle to balance these two relationships, further affecting modeling accuracy. Summary of the Invention

[0003] To address the problems of "pseudo-interaction" interference, global-local information conflict, and low modeling accuracy in existing models of human-vehicle lane-changing behavior, the present invention aims to provide a simulation model, method, and system for human-vehicle lane-changing behavior, enabling high-precision simulation and prediction of human-vehicle lane-changing behavior.

[0004] The technical solution adopted in this invention is as follows: The simulation model of lane-changing behavior of human drivers and vehicles includes the following process: Input module: used to acquire traffic flow data of the target area, including real-time status data of lane-changing vehicles, status data of surrounding vehicles of lane-changing vehicles, and road structure data; Input embedding and feature preprocessing module: used to perform dimensionality-upgrading preprocessing and dual processing of numerical embedding and location embedding on the traffic flow data to obtain the embedding feature matrix; LSTM short-time dynamic feature extraction module: used to extract short-time dynamic features of the traffic flow data and obtain a short-time dynamic feature matrix; Sparse feature optimization module: used to perform adaptive dense self-attention and sparse self-attention processing on the short-time dynamic feature matrix to obtain the feature optimization matrix; The encoder-decoder global integration module is used to obtain the global dependency of lane-changing behavior by utilizing the embedded feature matrix and the feature optimization matrix, and to obtain the output matrix. Output module: Used to obtain the simulation results of lane change behavior based on the output matrix.

[0005] Furthermore, the input embedding and feature preprocessing module includes: Feature preprocessing submodule: used to superimpose traffic flow data into a matrix form to obtain the input matrix; Numerical embedding submodule: used to map the input matrix to a high-dimensional space through a single convolution operation to obtain a numerical embedding matrix; The position embedding submodule is used to construct position codes by performing sine and cosine functions on the numerical embedding matrix to obtain the position embedding matrix. Feature fusion submodule: used to superimpose the numerical embedding matrix and the positional embedding matrix to obtain the embedded feature matrix.

[0006] Furthermore, the LSTM short-time dynamic feature extraction module includes a gating mechanism submodule and a temporal feature output submodule; The gating mechanism submodule controls information flow through an input gate, a forget gate, and an output gate. Its internal update formula is:

[0007]

[0008]

[0009]

[0010]

[0011]

[0012] In the formula, Learnable parameters: It is the sigmoid activation function. This is element-wise multiplication. These represent the input gate, forget gate, and output gate, respectively. For the input data at time step t, This is the hidden state from the previous time step. Candidate cell state, This represents the cell state at the previous time step. The temporal feature output submodule is used to output the short-time dynamic feature matrix after LSTM processing, as shown in the following formula:

[0013] In the formula, Represents the short-time dynamic feature matrix. This indicates the latent state and cell state after extraction. This represents the initial hidden state and the cell state. This represents traffic flow data.

[0014] Furthermore, the sparse feature optimization module performs adaptive dense self-attention and sparse self-attention processing on the short-term dynamic feature matrix through an adaptive sparse attention mechanism to obtain the feature optimization matrix, specifically: The multi-head query, key, and value tensors are generated through linear projection, as shown in the following formula:

[0015] In the formula, Represents the short-time dynamic feature matrix. , , For feature dimension, For time step, For batch size, The attention score is calculated using the following formula: (This refers to the number of attention heads.)

[0016] In the formula, Based on window size The relative position bias is used to optimize the attention weights using a two-branch mechanism, as shown in the following formula:

[0017]

[0018] In the formula, The attention score at time step t is the score at the zeroth step. Here is the attention score matrix at time step t. The attention score for the first step at time step t. This represents the row-by-row normalization function. Represents square ReLU function, Represents the attention matrix. Representing two normalized weights, the feature optimization matrix output by the adaptive sparse attention mechanism using this attention matrix is:

[0020] Furthermore, the encoder-decoder global integration module includes: Feature fusion submodule: used to add and fuse the embedded feature matrix and the feature optimization matrix to form a feature fusion matrix; Encoder submodule: used to perform multi-layer feature encoding on the feature fusion matrix, capture the global dependency of lane change behavior, and generate a global encoded feature matrix; Decoder submodule: Used for lane change state prediction decoding based on the global encoded feature matrix, mapping high-dimensional features into interpretable lane change parameters, which are used as the output matrix.

[0021] Furthermore, when performing multi-layer feature encoding on the feature fusion matrix to capture the global dependencies of lane-changing behavior and generate the global encoded feature matrix, the following formula is used:

[0022] In the formula, This refers to the Transformer encoder, which uses a multi-head self-attention mechanism and a feedforward neural network to model the input features, thereby capturing the complex patterns of a vehicle being affected by both global targets and local dynamics within the lane-change region. For the feature fusion matrix, for Feature representation after encoder processing For feature dimension, For time step, Batch size; When performing lane change state prediction decoding based on the global encoded feature matrix, mapping high-dimensional features into interpretable lane change parameters is done using the following formula:

[0023] In the formula, Indicates the decoder embedding layer, This represents the feature matrix output by the decoder embedding layer, which interacts with the encoder output to obtain the final output matrix. Represented as:

[0025] This invention also provides a method for simulating lane-changing behavior of human-driven vehicles. This method is applied to the lane-changing behavior simulation model of human-driven vehicles described above, and includes the following steps: Acquire traffic flow data for the target area, perform data calibration processing on the traffic flow data for the target area, and construct a traffic change dataset for people, vehicles, and drivers. Input the human-vehicle lane change dataset into the human-vehicle lane change behavior simulation model, train the model, and obtain the trained human-vehicle lane change behavior simulation model. The obtained prediction data is used to construct an input matrix using the sliding window method. The input matrix is ​​then input into the trained simulation model of human-vehicle lane-changing behavior. The simulation model outputs the longitudinal and lateral velocities of the lane-changing vehicle at the next moment.

[0026] Furthermore, the step of acquiring traffic flow data in the target area, performing data calibration processing on the traffic flow data in the target area, and constructing a traffic lane change dataset for pedestrians, vehicles, and cars includes: Data acquisition sub-steps: Extract the longitudinal and lateral positions and speeds of lane-changing vehicles, the longitudinal and lateral positions and speeds of the vehicle in front in the current lane and the vehicle to the left front and left rear in the target lane, the pressure values ​​of the lane in the current frame and the target lane in the current frame, and mark the missing information of vehicles in the target lane as -1. Missing value handling sub-step: The missing coordinate values ​​of the lane-changing vehicle are filled in using interpolation of the mean coordinates before and after the change of lanes, as shown in the following formula:

[0027]

[0028] In the formula, and This represents the position coordinates of the vehicle changing lanes at the current time t. and This represents the position coordinates of the vehicle that changed lanes at the previous time t. and This represents the position coordinates of the vehicle changing lanes at the next time step, from the current time t. Smoothing sub-step: A symmetric exponential moving average filtering algorithm is used to denoise the vehicle center point coordinates, speed, and lane change pressure. The formula is as follows:

[0029] In the formula, express The data values ​​after time smoothing i The index of the currently processed data point. Indicates time Traffic flow data at the time For exponential decay parameters, For symmetrical window length, This represents the total number of data points.

[0030] Furthermore, the input data is used to construct an input matrix using the sliding window method, which is then input into the trained simulation model of human-vehicle lane-changing behavior. The simulation model outputs the longitudinal and lateral velocities of the lane-changing vehicle at the next moment, specifically including the following process: Input matrix construction sub-step: before stacking The single-step state at each time step is given by the following formula:

[0031] In the formula, Indicates the length of the history window. For a moment The observation vector, For the first The first moment Values ​​of each dimension Represented as:

[0032] In the formula, This is the state matrix for lane-changing vehicles; This is the state matrix of the vehicle ahead in the current lane; The state matrix of the vehicle ahead in the target lane; The state matrix of vehicles following the target lane; This is a road structure data matrix; Model output sub-steps: The single-step output of the simulation model of human-driven vehicle lane-changing behavior is as follows:

[0033] In the formula, and These represent vehicles changing lanes at... The vertical and horizontal velocities of a moment, if focusing on several future moments... At each time step, the output is expanded to:

[0034] In the formula, each row express The velocity vector output at any given time.

[0035] This invention also provides a simulation system for lane-changing behavior of human-driven vehicles, including the following processes: Simulation scenario construction module: used to build a lane-changing environment road network in the simulation platform and generate traffic flow according to the negative displacement exponential distribution; Model loading module: used to integrate the pre-trained simulation model of human-driven vehicle lane-changing behavior described above into a virtual platform; Lane change behavior prediction module: This module uses data from the virtual platform to construct an input matrix through a sliding window, which is then input into the trained model. The model outputs longitudinal and lateral velocities and writes them back to the simulation platform to control the vehicle's lane change.

[0036] Compared with the prior art, the present invention has the following beneficial effects: This invention addresses the problems of "pseudo-interaction" interference, global-local information conflict, and low modeling accuracy in human-vehicle lane-changing behavior modeling. Based on a human-vehicle lane-changing dataset, it proposes an LA-Transformer lane-changing behavior simulation model (i.e., the human-vehicle lane-changing behavior simulation model) that integrates an adaptive sparse attention mechanism, a long short-term memory network, and a Transformer. The adaptive sparse attention mechanism filters key lane-changing features to eliminate "pseudo-interaction" noise, while the long short-term memory network's short-term dynamic modeling capability and the Transformer's global information integration capability achieve accurate characterization of human-vehicle lane-changing behavior. This invention solves the problems in existing lane-changing behavior modeling methods, such as "pseudo-interaction" interference in key feature judgment due to dense self-attention, the inability of fixed memory windows to balance global lane-changing pressure and local dynamic response, and insufficient modeling accuracy to meet high-fidelity simulation requirements. It provides reliable model support for highway traffic flow simulation, human-vehicle behavior analysis, and traffic safety assessment. Attached Figure Description

[0037] Figure 1 This is the network architecture diagram of the simulation model of human-driven vehicle lane-changing behavior (i.e., LA-Transformer) of this invention; Figure 2 This is a schematic diagram of the dataset extraction scenario in Embodiment 1 of the present invention (i.e., Freeway C); Figure 3 This is a comparison chart of six network loss values ​​in Embodiment 2 of the present invention; Figure 4 This is a comparison chart of the SpeedX performance of different models in Embodiment 2 of the present invention; Figure 5 This is a comparison chart of the SpeedY performance of different models in Embodiment 2 of the present invention; Figure 6(a) shows the frequency distribution and normal distribution of the output speed and the actual speed of the SpeedX dimensional model in Embodiment 2 of the present invention; Figure 6(b) shows the frequency distribution and normal distribution of the output speed and the actual speed of the SpeedY dimension model in Embodiment 2 of the present invention; Figure 7 This is the visualized road network map in Embodiment 3 of the present invention. Detailed Implementation

[0038] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, not all embodiments, and are not intended to limit the scope of the present invention. Furthermore, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion regarding the concepts disclosed in the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0039] To address the issues of "pseudo-interaction" interference, global-local information conflict, and low modeling accuracy in modeling human-vehicle lane-changing behavior, this invention proposes a method for constructing a simulation model of human-vehicle lane-changing behavior using the LA-Transformer model based on a human-vehicle lane-changing dataset. This method achieves high-precision simulation of lane-changing behavior through input embedding and feature preprocessing optimization, integration of long short-term memory network for short-term dynamic modeling, sparse optimization using an adaptive sparse attention mechanism, and global integration of encoder and decoder.

[0040] See Figure 1 The simulation model of lane-changing behavior of a human-driven vehicle in this invention includes: an input module, an input embedding and feature preprocessing module, an LSTM short-time dynamic feature extraction module, a sparse feature optimization module, an encoder-decoder global integration module, and an output module; wherein: The input module is used to receive traffic flow data of the target area, which includes real-time status data of lane-changing vehicles, status data of surrounding vehicles, and road structure data. The input embedding and feature preprocessing module is used to preprocess the input traffic flow data and then perform dual processing of numerical embedding and location embedding to enhance the expression of lane change behavior-related features and obtain an embedded feature matrix. Specifically, this input embedding and feature preprocessing module includes a feature preprocessing submodule, a numerical embedding submodule, a location embedding submodule, and a feature fusion submodule. Through the input embedding and feature preprocessing module, the input retains both numerical features and embeds temporal and location information, facilitating the subsequent extraction of global spatiotemporal features. The feature preprocessing submodule is used to process the input traffic flow data of pedestrians and vehicles changing lanes into a feature matrix form that can be recognized by the model. ;in, Represents the input matrix, Indicates batch size. This indicates the window length, and 15 indicates the input dimension.

[0041] The numerical embedding submodule performs a one-dimensional convolution operation to map the transposed input matrix to a higher dimension, thereby upscaling the low-dimensional features to a higher-dimensional space and generating a numerical embedding feature matrix; the specific formula is as follows:

[0042] In the formula, This represents the transposed input. This represents the convolution operation.

[0043] The location embedding submodule performs temporal location encoding on the numerical embedding feature matrix using sine and cosine functions, preserving the temporal series correlation of the input features, and generating a location embedding feature matrix; the specific formula is as follows:

[0044] In the formula, For time step index, For dimensional indexing, Indicates the window length.

[0045] The feature fusion submodule superimposes the numerical embedding matrix and the positional embedding matrix to obtain the final embedded feature matrix, as shown in the following formula:

[0046] In the formula, Indicates the input dimension. This represents the feature matrix after numerical embedding. This represents the feature matrix after position encoding. This represents the feature matrix after double embedding.

[0047] This invention introduces a dual embedding operation by constructing an input embedding and feature preprocessing module, and uses a single-layer convolution operation to achieve feature dimensionality enhancement, which can strengthen the correlation between the interaction features of lane-changing vehicles and surrounding vehicles and the road environment features; and retains the temporal logic of traffic flow data through sine and cosine position encoding, avoiding the confusion of lane-changing behavior logic caused by the lack of temporal awareness in the Transformer architecture, improving the model's ability to capture the temporal continuity of the lane-changing process, while ensuring the expression efficiency and discriminability of high-dimensional features.

[0048] The LSTM short-time dynamic feature extraction module is used to extract short-time dynamic features of traffic flow data through a gating mechanism to obtain a short-time dynamic feature matrix. Specifically, the LSTM short-time dynamic feature extraction module includes a gating mechanism submodule and a temporal feature output submodule. After the embedded features are generated, the LSTM module is used to capture traffic flow data. The short-time dynamic characteristics of LSTM receiver input By extracting temporal dependencies through a multi-layer recurrent network, strong temporal support is provided for subsequent sparse optimization, wherein: The gating mechanism submodule is used to construct an information filtering mechanism through an input gate, a forget gate, and an output gate, selectively retaining and forgetting short-term dynamic information in the input; the internal update formula of the gating mechanism submodule is:

[0049]

[0050]

[0051]

[0052]

[0053]

[0054] In the formula, Learnable parameters: It is the sigmoid activation function. This is element-wise multiplication. These represent the input gate, forget gate, and output gate, respectively. For the input data at time step t, This is the hidden state from the previous time step. Candidate cell state, This represents the cell state at the previous time step.

[0055] The temporal feature output submodule is used to calculate and extract short-time dynamic features during lane changing based on cell state and output gate, and generate a short-time dynamic feature matrix. The specific formula is as follows:

[0056] In the formula, Representing temporal characteristics, This indicates the latent state and cell state after extraction. This represents the initial hidden state and the cell state. This indicates input.

[0057] This invention introduces a gated temporal modeling mechanism through an LSTM short-time dynamic feature extraction module to achieve accurate screening of short-time dynamic information through multi-gating collaboration, thereby enhancing the module's ability to capture short-time sudden features of lane changes. It also compensates for the shortcomings of Transformer in short-time rapid behavior modeling, especially in scenarios with strong randomness in human and vehicle behavior, and improves the extraction accuracy of short-time dynamic features of lane changes.

[0058] The sparse feature optimization module is used to retain effective interaction features from the short-term dynamic feature matrix through dense self-attention branches and filter out noise through sparse self-attention branches, and then combine adaptive weight fusion to filter key lane-changing features to obtain the feature optimization matrix; this sparse feature optimization module is designed to solve... The "pseudo-interaction" problem caused by dynamic sparsity in dynamic sparsity leads to the introduction of the ASSA module (Adaptive Sparse Self-Attention Module) for further optimization. ASSA generates multi-head query, key, and value tensors through linear projection, as shown in the following formula:

[0059] In the formula, Represents the short-time dynamic feature matrix. , , For feature dimension, For time step, For batch size, The attention score is calculated using the following formula: (This refers to the number of attention heads.)

[0060] In the formula, Based on window size The relative positional bias. ASSA uses a two-branch mechanism to optimize attention weights, as shown in the following formula:

[0061]

[0062] In the formula, The attention score at time step t is the score at the zeroth step. Here is the attention score matrix at time step t. The attention score for the first step at time step t. This represents the row-by-row normalization function. Represents square ReLU function, Represents the attention matrix. This represents two normalized weights. Using this attention matrix, the ASSA module outputs:

[0063] The ASSA module enables the model to focus on key features, adaptively filter information closely related to lane-changing behavior, and effectively suppress unnecessary interactions caused by missing data and noise.

[0064] Specifically, the adaptive sparse attention module includes a dense self-attention submodule, a sparse self-attention submodule, and an adaptive weight fusion submodule, wherein: The dense self-attention submodule is used to perform dense attention calculation on short-time dynamic feature matrix pairs, capture potential interaction features related to lane change, and generate a dense interaction feature matrix. The sparse self-attention submodule is used to perform local attention filtering on the short-time dynamic feature matrix, filter out irrelevant noise, and generate a sparse and effective feature matrix. The adaptive weight fusion submodule is used to dynamically adjust the fusion weights of the two types of feature matrices based on the lane change scenario. It obtains the feature optimization matrix by weighted summation to filter key lane change features.

[0065] This invention employs a sparse feature optimization module to fuse a dual-branch attention mechanism, thereby accurately extracting the core features of lane changes by drivers and vehicles through a branching process of "dense capture-sparse filtering-adaptive fusion". This avoids the redundant overhead of traditional Transformer attention calculations and further improves the model's ability to learn and model key interactive relationships in complex traffic flows.

[0066] The encoder-decoder global integration module is used to simultaneously utilize the embedded feature matrix and the feature optimization matrix to capture the global dependency of lane-changing behavior, achieving a global mapping from input features to output motion state and obtaining the output matrix. To simultaneously utilize the embedded global features and the local dynamic information captured by LSTM, this module fuses the data embedding and the ASSA output. Specifically, the encoder-decoder global integration module includes a feature fusion submodule, an encoder submodule, and a decoder submodule, wherein: The feature fusion submodule is used to fuse embedded features and adaptive sparse attention-optimized features, that is, to add the embedded feature matrix and the feature optimization matrix and fuse them to form a feature fusion matrix. The specific calculation formula is as follows:

[0067] In the formula, This represents the fused global feature matrix. This fusion not only compensates for the shortcomings of LSTM in capturing global information, but also provides richer spatiotemporal information for the subsequent Transformer encoder.

[0068] The encoder submodule is used to perform multi-layer feature encoding on the feature fusion matrix, capture the global dependencies of lane-changing behavior, and generate a global encoded feature matrix; specifically, it combines features. As input to the Transformer encoder, global dependencies are captured through multiple layers of encoders, represented as:

[0069] In the formula, This refers to the Transformer encoder, which uses a multi-head self-attention mechanism and a feedforward neural network to model the input features, thereby capturing the complex patterns of a vehicle being affected by both global targets and local dynamics within the lane-change region. For the feature fusion matrix, for Feature representation after encoder processing For feature dimension, For time step, This refers to the batch size.

[0070] The decoder submodule is used for lane change state prediction decoding based on the global encoded feature matrix, mapping high-dimensional features into interpretable lane change parameters, i.e., a speed prediction matrix. Specifically, the decoder processes sequences... It is generated through the embedding layer, and the formula is as follows:

[0071] In the formula, Indicates the decoder embedding layer, The feature matrix output by the decoder embedding layer interacts with the encoder output to obtain the final output matrix, which can be represented as:

[0072] This invention introduces a global modeling mechanism through an encoder-decoder global integration module, achieving accurate mapping of lane-changing behavior from local features to global states. This further enhances the model's simulation capability of the overall lane-changing process, ensuring the reproduction of lane-changing behavior of drivers and vehicles in different scenarios. It also demonstrates better lane-changing state prediction accuracy and behavior simulation stability. This module not only optimizes the model's lane-changing simulation effect in real road scenarios but also improves the model's adaptability and robustness to different road environments, significantly enhancing the lane-changing behavior simulation capability of the LA-Transformer model.

[0073] The output module is used to output the speed prediction matrix as predicted speed, specifically the longitudinal and lateral speeds of the lane-changing vehicle at the next moment or the speeds at the next S time steps, to obtain the simulation results of the lane-changing behavior.

[0074] In the above scheme, the present invention improves the Transformer model by introducing an input embedding and feature preprocessing module, a sparse feature optimization module, and an LSTM short-time dynamic feature extraction module, thereby further improving the model's performance, enhancing its ability to utilize multi-source traffic flow data in lane-changing scenarios involving people, vehicles, and drivers, improving the accuracy of capturing key features of lane-changing behavior in complex road environments, and reducing lane-changing state prediction bias and feature noise interference.

[0075] This invention provides a method for simulating lane-changing behavior of human-driven vehicles, comprising the following processes: Traffic flow data within the target area is acquired and data calibration is performed. The traffic flow data includes real-time status data of lane-changing vehicles, road structure data, and surrounding vehicle status data. The data calibration process includes missing value processing and smoothing processing to obtain a human-vehicle lane-changing dataset. A simulation model of human-driver lane changing behavior is constructed using LA-Transformer. The LA-Transformer lane changing behavior simulation model is obtained by input embedding and feature preprocessing of Transformer architecture, short-term dynamic modeling and integration of long short-term memory network, sparse optimization of adaptive sparse attention mechanism and global integration of encoder and decoder. The LA-Transformer lane change behavior model was trained using a human-vehicle lane change dataset to obtain a trained LA-Transformer human-vehicle lane change behavior model. The real-time status data of the lane-changing vehicle, road structure data, and surrounding vehicle status data at each time step within a predetermined time period are used as input data to form an input matrix, which is then input into the trained LA-Transformer model. The trained LA-Transformer model outputs the motion state of the lane-changing vehicle at the next time step.

[0076] Furthermore, the acquisition of traffic flow data in the target area, followed by data calibration processing to construct a lane change dataset for vehicles and drivers, includes: Data acquisition sub-steps: Extract lane-changing vehicle data from the target area, namely the longitudinal and lateral positions and speeds of lane-changing vehicles; surrounding vehicle data, namely the longitudinal and lateral positions and speeds of the vehicle in front of the current lane and the vehicle to the left front and left rear of the target lane; and road structure data, namely the lane in the current frame and the pressure value of the target lane in the current frame, and mark the missing information of vehicles in the target lane as -1. Missing value handling sub-step: Missing values ​​of the lane-changing vehicle's position coordinates are filled by interpolation using the average coordinates before and after the change of lanes, as shown in the following formula:

[0077]

[0078] In the formula, and This indicates the current position coordinates of the vehicle changing lanes.

[0079] Smoothing sub-step: A symmetric exponential moving average filtering algorithm is used to denoise the vehicle center point coordinates, speed, and lane change pressure. The formula is as follows:

[0080] In the formula, express The data values ​​after time smoothing i The index of the currently processed data point. Indicates time Data at that time, For exponential decay parameters, For symmetrical window length, This represents the total number of data points.

[0081] Furthermore, the input data is used to construct an input matrix using the sliding window method, which is then input into the trained model. The model outputs the longitudinal and lateral velocities of the lane-changing vehicle at the next moment, including: Input matrix construction sub-step: before stacking The single-step state at each time step is given by the following formula:

[0082] In the formula, Indicates the length of the history window. For a moment The observation vector, For the first The first moment The values ​​for each dimension. To further demonstrate the structured nature of the input data, The decomposition into a block matrix form of feature groups highlights the evolution of each sub-vector over time. Specifically, It can be represented as:

[0083] In the formula, This is the state matrix for lane-changing vehicles; This is the state matrix of the vehicle ahead in the current lane; The state matrix of the vehicle ahead in the target lane; The state matrix of vehicles following the target lane; This is a road structure data matrix.

[0084] Model output sub-step: The goal of the model is to output the vehicle's motion state at future time steps. Taking a single-step output as an example, the output focuses on the velocity vector at the next time step, and the output is defined as:

[0085] In the formula, and These represent vehicles changing lanes at... The vertical and horizontal velocities of a given moment. If we focus on several future moments... At each time step, the output is expanded to:

[0086] In the formula, each row express The velocity vector output at any given time.

[0087] This invention provides a simulation system for lane-changing behavior of human-driven vehicles, comprising: Simulation scenario construction module: used to build a virtual lane-changing environment road network in the simulation platform, configure detectors to collect data, and generate traffic flow according to the negative displacement distribution; Model loading module: Used to integrate pre-trained LA-Transformer simulation models of human-vehicle lane-changing behavior into the platform; Lane change behavior prediction module: It is used to collect the vehicle's own state, the state of surrounding vehicles and road data in the previous T time steps in the simulation in real time, and input them into the trained LA-Transformer model. After the model outputs the longitudinal and lateral speeds of the lane-changing vehicle at the next moment, it writes the speed command back to the simulation platform through the interface to control the vehicle to perform lane change behavior.

[0088] Example 1 This embodiment addresses the issues of "pseudo-interaction" noise interference and conflicts between global lane change constraints and local dynamic responses during lane changes by human-driven vehicles. It constructs an LA-Transformer simulation model for human-driven vehicle lane change behavior. The core architecture is described in [link to core architecture]. Figure 1 It includes an input module, an input embedding and feature preprocessing module, an LSTM short-time dynamic feature extraction module, a sparse feature optimization module, an encoder-decoder global integration module, and an output module. The specific implementation is as follows: Input module: Responsible for receiving traffic flow data. The input data consists of 15-dimensional features, specifically including: the longitudinal and lateral positions and speeds of vehicles changing lanes, the longitudinal and lateral positions and speeds of vehicles in front in the current lane, the longitudinal and lateral positions and speeds of vehicles in front in the target lane, the longitudinal and lateral positions and speeds of vehicles behind in the target lane, the current lane markings, and lane change pressure values.

[0089] Input embedding and feature preprocessing module: Transforms 15-dimensional features into high-dimensional temporal features in four steps: Feature preprocessing submodule: Superimposes 15-dimensional data (i.e., 15-dimensional traffic flow data) into a matrix form; obtains the input matrix; The numerical embedding submodule maps the 15-dimensional input features to a 128-dimensional high-dimensional space through a single convolution operation, resulting in a numerical embedding matrix.

[0090] The position embedding submodule constructs position codes by applying a sine-cosine function to the numerical embedding matrix, thus obtaining the position embedding matrix. Feature fusion submodule: Superimposes the numerical embedding matrix and the positional embedding matrix to obtain the final embedded feature matrix.

[0091] LSTM Short-Term Dynamic Feature Extraction Module: Simultaneously sets up a 2-layer Long Short-Term Memory network structure, with the hidden layer dimension and the embedding dimension being consistent. After processing by the module, the input matrix is ​​output as a matrix containing short-term dynamic features.

[0092] Sparse Feature Optimization Module: An adaptive sparse attention mechanism is introduced to adaptively process the short-term dynamic feature matrix with dense self-attention and sparse self-attention, and finally output the optimized lane change key feature matrix.

[0093] Encoder-decoder global integration module: Achieves accurate mapping of lane-changing behavior from local features to global state, executed in three steps: Feature fusion submodule: Overlays the embedded feature matrix and the feature optimization matrix to integrate temporal dynamics and key interaction information; Encoder submodule: Captures global dependencies through 6 layers of encoders, such as the constraint of the acceleration lane end point on lane change timing and the impact of main road traffic flow on lane change gaps; Decoder submodule: First, it performs embedding processing on the output sequence, and then interacts with the encoder output through a 6-layer decoder to map global and local features into interpretable lane change parameters.

[0094] Output module: Through fully connected layers and activation functions, interpretable lane change parameters are mapped and output as longitudinal and lateral velocities for the next 10 time steps, with the velocity matrix dimension being "10×2".

[0095] The main parameters and their specific settings in this embodiment are shown in Table 1.

[0096] Table 1

[0097] Example 2 This embodiment provides a method for simulating lane-changing behavior of human-driven vehicles. This method is applied to the LA-Transformer lane-changing behavior simulation model of human-driven vehicles provided in the above embodiment, and includes the following steps: Filter the lane change trajectory data of pedestrians, vehicles, and drivers in the merging zone of the CitySim public dataset Freeway C scenario, see [link / reference]. Figure 2 The selection criteria were "lane change starting in the acceleration lane, ending in the main lane, and only one lane change." The longitudinal and lateral positions and speeds of the lane-changing vehicle, as well as the status information of surrounding vehicles (i.e., the vehicle in front in the current lane and the vehicles in front and behind in the target lane), were obtained to form a dataset of 872 valid lane-change trajectories. The data was preprocessed: longitudinal and lateral speeds were calculated using the central difference method, and speed noise was filtered out using a symmetric exponential moving average filter; for cases where information about vehicles in front and behind in the target lane was missing, it was marked as -1 and filled in using interpolation at adjacent time steps.

[0098] The X-axis represents the longitudinal direction of vehicle travel, and the Y-axis represents the lateral direction of vehicle travel. The extracted actual lane-change trajectory can be represented as a set. , of which Lane change trajectory , The sequence length of the lane change trajectory is set to 10. Each... It contains 15 features, see Table 2, and the model output is the future. =Latitude and lateral velocities over 10 time steps.

[0099] Table 2

[0100] The preprocessed lane-change trajectory is used as input to the LA-Transformer. To evaluate the model's ability to model lane-change behavior of drivers and vehicles, the superiority of the proposed method will be verified from the perspectives of model convergence performance comparison, comprehensive performance index evaluation, single lane-change scenario effect verification, and generalization performance verification on a complete dataset.

[0101] (1) Comparison of model convergence performance During the model training phase, the convergence speed and stability of each model are evaluated by recording the loss changes during the training process. Figure 3 The loss curves of LSTM, Transformer, Seq2Seq, BiLSTM, Autoformer, and the proposed LA-Transformer during training were compared. Experimental results show that LA-Transformer has advantages in both final loss convergence and training stability. Specifically, LA-Transformer's loss decreases rapidly in the early stages of training, exhibiting a clear convergence trend and ultimately reaching the lowest convergence level among all models. In contrast, the Seq2Seq model's loss value remains at a high level at the end of training, demonstrating a large convergence error. Furthermore, the loss curve of LA-Transformer exhibits less overall fluctuation, which is better than that of LSTM and Transformer models. These results indicate that LA-Transformer possesses stronger convergence ability and better training stability in modeling lane-changing behavior of human-driven vehicles.

[0102] (2) Comprehensive performance index evaluation To quantitatively evaluate the performance of each model across different dimensions, the longitudinal speed (SpeedX), lateral speed (SpeedY), and the total average of all features (Total) of the lane-changing vehicle were calculated in terms of MAE, RMSE, and other metrics. The performance of ADE and FDE indicators was analyzed, and the results are summarized in Table 3.

[0103] Table 3

[0104] Experiments show that the proposed LA-Transformer outperforms other comparative models across all dimensions, demonstrating a certain performance advantage. Taking the SpeedX dimension as an example, the LA-Transformer model achieves a MAE of 0.0261 and an RMSE of 0.0336. The performance was excellent, achieving an ADE of 0.9882, an ADE of 0.0261, and an FDE of 0.0315. Compared to the suboptimal Autoformer model, the LA-Transformer reduced its MAE by 0.0145 and its RMSE by 0.0166. The accuracy improved by 3.07%, ADE decreased by 0.0145, and FDE decreased by 0.0211. In both SpeedY and Total dimensions, LA-Transformer showed similar trends, demonstrating high accuracy and stability, outperforming other comparative models and indicating its superiority in modeling human-vehicle lane-changing behavior.

[0105] (3) Verification of the effect of a single lane change scenario To visually demonstrate the modeling capabilities of different models for lane-changing behavior of drivers and vehicles in specific scenarios, a lane-changing segment with vehicle number 75 was randomly selected from the CitySim merging zone dataset. The learning performance of six models—LSTM, Transformer, Seq2Seq, BiLSTM, Autoformer, and LA-Transformer—on this lane-changing segment was analyzed in terms of longitudinal speed (SpeedX) and lateral speed (SpeedY).

[0106] Figure 4 The output of each model for SpeedX is shown. Due to the large vertical velocity, the models need to have strong dynamic capture capabilities. Overall, LA-Transformer performs excellently across the entire data segment, with its output velocity curve closely matching the real velocity, even in regions of rapid velocity changes. LSTM and Seq2Seq, on the other hand, are relatively weaker in depicting local details. Further observation of a magnified local area reveals that LA-Transformer can accurately capture even slight velocity fluctuations, effectively reproducing the true trend of SpeedX changes in this data segment.

[0107] Figure 5The output of each model for SpeedY is shown. Compared to SpeedX, SpeedY exhibits smaller overall fluctuations, requiring the model to possess stronger detail capture capabilities. Overall, LA-Transformer effectively characterizes changes in lateral velocity, especially in areas with slight velocity fluctuations, responding quickly and outputting accurately. In contrast, LSTM and Seq2Seq are significantly less detailed, exhibiting distortion in local data. Zooming in on local data further validates LA-Transformer's advantage in capturing velocity changes, demonstrating a significantly superior ability to characterize details compared to other models. In summary, LA-Transformer performs stably in scenarios with high vertical velocity and accurately captures changes in scenarios with low but subtle lateral velocity fluctuations, showcasing strong learning capabilities.

[0108] (4) Validation of generalization performance on the complete dataset To verify the proposed LA-Transformer model's adaptability to the heterogeneity of different driver behaviors, as well as its consistency and generalization ability across the entire dataset, the frequency distributions of the model's output speed and actual speed for longitudinal speed (SpeedX) and lateral speed (SpeedY) for all lane-change segments were plotted on the complete dataset, and normal distribution curves were drawn to analyze their statistical characteristics. Overall, the output speed distribution of the LA-Transformer model is highly consistent with the actual speed distribution in terms of statistical characteristics, with small errors, demonstrating the model's robustness under diverse driving scenarios.

[0109] Specifically, in the SpeedX dimension, as shown in Figure 6(a), the mean of the actual speed distribution is 20.0818, and the standard deviation is 2.7934; while the mean of the model output speed is 20.0697, and the standard deviation is 2.8294. The comparison shows that the absolute error of the mean is 0.0121, the relative absolute error is 0.0603%, the absolute error of the standard deviation is 0.0360, and the relative absolute error is 1.2888%. In the SpeedY dimension, as shown in Figure 6(b), the mean of the actual speed distribution is 0.4780, and the standard deviation is 0.3446; while the mean of the model output speed is 0.4688, and the standard deviation is 0.3434. The comparison shows that the absolute error of the mean is 0.0092, the relative absolute error is 1.9247%, the absolute error of the standard deviation is 0.0012, and the relative absolute error is 0.3482%. The above results show that the model output speed distribution is highly consistent with the real distribution in terms of mean and standard deviation, and the distribution is balanced in both high-speed and low-speed regions without systematic bias. This verifies the robustness of LA-Transformer in diverse driving scenarios and demonstrates its ability to accurately characterize the behavioral characteristics of different drivers.

[0110] Example 3 This embodiment provides a simulation system for human-vehicle lane-changing behavior based on LA-Transformer, used to implement the human-vehicle lane-changing behavior modeling method described in Embodiment 2 using the LA-Transformer lane-changing behavior simulation model described in Embodiment 1 above, including the following steps: The first step is to initiate the simulation scenario construction process, building a lane-changing virtual scenario based on the SUMO microscopic traffic simulation platform. This scenario is built using NETEDIT within SUMO, and the constructed road network scenario includes 5 nodes and 4 edges. The specific parameters of the nodes and edges are shown in Tables 4 and 5, respectively. The road network mainly consists of a single-lane main lane and a single-lane ramp (including an acceleration lane). The visualized road network is shown below. Figure 7 As shown. The main line is 600m long, with 200m each upstream, midstream and downstream of the merging zone, where the speed limit is 33.33m / s (approximately 120km / h). The ramps are approximately 200m long, with a speed limit of 11.11m / s (approximately 40km / h). An acceleration lane is also provided in the merging zone, which is 200m long and 3.5m wide.

[0111] Table 4

[0112] Table 5

[0113] In the SUMO road network, detectors can be deployed to collect traffic flow data. The detectors used in this experiment include two types: (1) Induction Loops Detectors (E1): As a single-point fixed-point detection device, its function is similar to that of a loop detector in reality. It can count the total number of vehicles passing through the detection point, traffic flow and average speed and other basic parameters within a specific time period; (2) Lanearea Detectors (E2): As a continuous area detection device, its function is similar to that of a video surveillance system. It can obtain dynamic information such as vehicle congestion index, parking time and delay time within the coverage area.

[0114] In this embodiment, the detectors are deployed in a composite layout: three E1 detectors are placed at equal intervals of 20m along the centerline of the main road in the merging zone, and four E1 detectors are placed 20m upstream and downstream of the merging zone entrance and exit. The sampling period for all E1 detectors is uniformly set to 30s. To enhance the full-process monitoring capability, an array of E2 detectors is continuously deployed along the main road to form a detection network covering the entire single-lane road section, enabling dynamic tracking of vehicles entering, passing through, and leaving the merging zone. The specific configuration information of the detectors is shown in Table 6.

[0115] Table 6

[0116] The second step involves loading the model, integrating the trained LA-Transformer human-vehicle lane-changing behavior model into the simulation system. Model loading strictly adheres to the "LA-Transformer model with adaptive sparse attention mechanism" architecture, ensuring that the model input consists of 15-dimensional features and the output is the longitudinal and lateral velocity matrices for the next 10 time steps. The interaction between the model and the simulation platform is completed via SUMO's TraCI interface, disabling SUMO's default LC2013 lane-changing model to ensure that the LA-Transformer model completely takes over the lane-changing decisions and behavioral control of the human-vehicle.

[0117] The third step involves lane change prediction and vehicle control. A sliding window method is used to construct the model input matrix and drive the simulation. The 15-dimensional features from the first 10 time steps are stacked using the window length to form the input matrix. This input matrix is ​​then fed into the LA-Transformer model. The model undergoes input embedding, LSTM short-time dynamic capture, ASSA sparse attention optimization, and encoder-decoder global integration to output the longitudinal and lateral velocities of the driver and vehicle for the next 10 time steps. The velocity results are written back to the SUMO simulation environment via the TraCI interface, and commands are invoked to control the driver and vehicle to perform lane change behavior at the predicted speed, ensuring a smooth transition of longitudinal and lateral velocities during lane changes.

[0118] This embodiment provides a simulation system for lane-changing behavior of human-vehicles based on LA-Transformer. The system has a simple structure and introduces LSTM short-time dynamic modeling, adaptive sparse attention, encoder-decoder global integration and linear projection output techniques to characterize lane-changing behavior, thereby achieving high-precision lane-changing behavior simulation. It can quickly and accurately depict the lane-changing process of human-vehicles merging from the acceleration lane into the main lane.

[0119] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A simulation model of human lane-changing behavior in a vehicle, characterized in that, The method comprises the following steps: An input module is configured to obtain traffic flow data of a target area, wherein the traffic flow data comprises real-time state data of a lane-changing vehicle, state data of surrounding vehicles of the lane-changing vehicle, and road structure data; An input embedding and feature preprocessing module is configured to perform dual processing of numerical embedding and position embedding on the traffic flow data to obtain an embedded feature matrix; An LSTM short-time dynamic feature extraction module is configured to extract short-time dynamic features of the traffic flow data to obtain a short-time dynamic feature matrix; A sparse feature optimization module is configured to perform adaptive dense self-attention and sparse self-attention processing on the short-time dynamic feature matrix to obtain a feature optimization matrix; An encoder-decoder global integration module is configured to obtain a global dependence relationship of a lane-changing behavior by using the embedded feature matrix and the feature optimization matrix to obtain an output matrix; An output module is configured to obtain a lane-changing behavior simulation result according to the output matrix.

2. The human lane-changing behavior simulation model according to claim 1, wherein, The input embedding and feature preprocessing module comprises: A feature preprocessing submodule is configured to stack the traffic flow data into a matrix form to obtain an input matrix; A numerical embedding submodule is configured to map the input matrix to a high-dimensional space by performing one-layer convolutional operation on the input matrix to obtain a numerical embedding matrix; A position embedding submodule is configured to construct position encoding by using a sine-cosine function on the numerical embedding matrix to obtain a position embedding matrix; A feature fusion submodule is configured to stack the numerical embedding matrix and the position embedding matrix to obtain the embedded feature matrix.

3. The human lane-changing behavior simulation model according to claim 1, wherein, The LSTM short-time dynamic feature extraction module comprises a gating mechanism submodule and a time sequence feature output submodule; The gating mechanism submodule controls information flow by using an input gate, a forgetting gate and an output gate, and the internal update formula is as follows: wherein, are learnable parameters: is a sigmoid activation function, is an element-wise multiplication, respectively represent the input gate, the forget gate and the output gate, is the input data at time step t, is the hidden state of the previous time step, is the candidate cell state, is the cell state of the previous time step; The time sequence feature output submodule is configured to output a short-time dynamic feature matrix processed by the LSTM, and the formula is as follows: In the formula, represents a short-time dynamic feature matrix, represents the extracted hidden state and cell state, represents the initial hidden state and cell state, represents traffic flow data.

4. The human lane-changing behavior simulation model of claim 1, wherein, The sparse feature optimization module performs adaptive dense self-attention and sparse self-attention processing on the short-time dynamic feature matrix by using an adaptive sparse attention mechanism to obtain a feature optimization matrix, and the specific process is as follows: Multi-head queries, key and value tensors are generated by linear projection, and the formula is as follows: In the formula, represents a short-time dynamic feature matrix, , , is a feature dimension, is a time step, is a batch size, is the number of attention heads, and the attention score calculation formula is: In the formula, is the relative position bias based on the window size The double-branch mechanism is used to optimize the attention weight, and the formula is as follows: wherein, is the attention score at the zeroth step on time step t, is the attention score matrix on time step t, is the attention score at the first step on time step t, denotes a row-wise normalization function, denotes a square function, denotes an attention matrix, denotes two normalization weights, with which the feature optimization matrix output by the adaptive sparse attention mechanism is: 。 5. The human lane-changing behavior simulation model according to claim 1, wherein, The encoder-decoder global integration module comprises: A feature fusion submodule is configured to add and fuse the embedded feature matrix and the feature optimization matrix to form a feature fusion matrix; An encoder submodule is configured to perform multi-layer feature encoding on the feature fusion matrix to capture a global dependence relationship of the lane-changing behavior and generate a global encoding feature matrix; A decoder submodule is configured to perform lane-changing state prediction decoding based on the global encoding feature matrix, map high-dimensional features into interpretable lane-changing parameters, and use the lane-changing parameters as the output matrix.

6. The lane-changing behavior simulation model of the human-driven vehicle according to claim 5, wherein: When the feature fusion matrix is subjected to multi-layer feature encoding to capture a global dependence relationship of the lane-changing behavior and generate a global encoding feature matrix, the following formula is used: wherein, denotes a Transformer encoder, which models the input features using multi-head self-attention mechanism and feed-forward neural networks, thereby capturing the complex patterns of the vehicle being influenced by both global targets and local dynamics within the lane-changing region, is a feature fusion matrix, is is a feature representation after encoder processing, is a feature dimension, is a time step, is a batch size; When lane-changing state prediction decoding is performed based on the global encoding feature matrix to map high-dimensional features into interpretable lane-changing parameters, the following formula is used: wherein represents a decoder embedding layer, represents a decoder embedding layer output feature matrix, which interacts with the encoder output to obtain a final output matrix, the output matrix is represented as: 。 7. A method of simulating lane changing behavior of human-driven vehicles, characterized by, This method, applied to the simulation model of lane-changing behavior of human-driven vehicles as described in any one of claims 1-6, includes the following steps: Acquire traffic flow data for the target area, perform data calibration processing on the traffic flow data for the target area, and construct a traffic change dataset for people, vehicles, and drivers. Input the human-vehicle lane change dataset into the human-vehicle lane change behavior simulation model, train the model, and obtain the trained human-vehicle lane change behavior simulation model. The obtained prediction data is used to construct an input matrix using the sliding window method. The input matrix is ​​then input into the trained simulation model of human-vehicle lane-changing behavior. The simulation model outputs the longitudinal and lateral velocities of the lane-changing vehicle at the next moment.

8. The human vehicle lane change behavior simulation method of claim 7, wherein, The process of acquiring traffic flow data in the target area, performing data calibration on the traffic flow data in the target area, and constructing a traffic change dataset for pedestrians, vehicles, and cars includes: Data acquisition sub-steps: Extract the longitudinal and lateral positions and speeds of lane-changing vehicles, the longitudinal and lateral positions and speeds of the vehicle in front in the current lane and the vehicle to the left front and left rear in the target lane, the pressure values ​​of the lane in the current frame and the target lane in the current frame, and mark the missing information of vehicles in the target lane as -1. Missing value handling sub-step: The missing coordinate values ​​of the lane-changing vehicle are filled in using interpolation of the mean coordinates before and after the change of lanes, as shown in the following formula: In the formula, and represents the position coordinate of the lane-changing vehicle at the current time t, and represents the position coordinate of the lane-changing vehicle at the previous time of the current time t, and represents the position coordinate of the lane-changing vehicle at the next time of the current time t. Smoothing sub-step: A symmetric exponential moving average filtering algorithm is used to denoise the vehicle center point coordinates, speed, and lane change pressure. The formula is as follows: In the formula, express The data values ​​after time smoothing i Index of the data point currently being processed. Indicates time Traffic flow data at the time For exponential decay parameters, For symmetrical window length, This represents the total number of data points.

9. The human vehicle lane change behavior simulation method of claim 7, wherein, The input data is constructed using the sliding window method to form an input matrix, which is then input into the trained simulation model of human-vehicle lane-changing behavior. The simulation model outputs the longitudinal and lateral velocities of the lane-changing vehicle at the next moment, specifically including the following process: Input matrix construction sub-step: pre-stack The single-step state of one time step is formulated as follows: wherein denotes the history window length, is the observation vector at time is the value of the dimension at time is the value of the dimension at time is denoted by: wherein is a lane-changing vehicle state matrix; is a current lane front vehicle state matrix; is a target lane front vehicle state matrix; is a target lane rear vehicle state matrix; is a road structure data matrix; Model output sub-steps: The single-step output of the simulation model of human-driven vehicle lane-changing behavior is as follows: where and denote the longitudinal and lateral speed of the changing lane vehicle at time instant, if focusing on a number of future time steps, the output extends to: and and where each row represents the velocity vector output at that time instant.

10. A human-driven vehicle lane change behavior simulation system, characterized by, The process includes the following: Simulation scenario construction module: used to build a lane-changing environment road network in the simulation platform and generate traffic flow according to the negative displacement exponential distribution; Model loading module: used to integrate the trained simulation model of lane-changing behavior of human-driven vehicles as described in claim 7 on a virtual platform; Lane change behavior prediction module: This module uses data from the virtual platform to construct an input matrix through a sliding window, which is then input into the trained model. The model outputs longitudinal and lateral velocities and writes them back to the simulation platform to control the vehicle's lane change.

Images