Lane changing behavior prediction method, system, equipment and medium

Abstract

The invention discloses a lane changing behavior prediction method, system and device and a medium, and relates to the technical field of vehicle track prediction. Comprising the following steps: extracting vehicle feature information of a lane changing vehicle with a lane changing behavior and surrounding vehicles from a HighD vehicle trajectory data set; the vehicle feature information is processed, and a processing result serves as a key input variable; taking a double-layer convolutional neural network 2-CNN as a lane change prediction model, and training the lane change prediction model through the key input variables to obtain a trained lane change prediction model; and in a lane changing prediction process, inputting data extracted from image data of a vehicle to be subjected to lane changing and surrounding vehicles into the trained lane changing prediction model for lane changing behavior prediction, and obtaining a lane changing behavior prediction result. According to the method, the accuracy of lane changing behavior prediction is remarkably improved, and powerful support is provided for optimization and application of an intelligent traffic technology.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle trajectory prediction, and particularly relates to a lane change behavior prediction method, system, device, and medium. Background Art

[0002] In the modern traffic system, the vehicle lane change behavior, as a basic operation during driving, widely exists in various road scenarios. The lane change behavior is not simply a vehicle displacement, but involves complex interactions of vehicles in a two-dimensional space. It not only requires speed adjustment but also is accompanied by dynamic changes in lateral displacement, which makes the interaction between vehicles more complex and has a profound impact on road traffic safety. According to the data of the "Annual Report on Traffic Safety Facts" of the United States, in 2020, there were 432,427 motor vehicle collision traffic accidents caused by lane change and merging operations, accounting for 5.3%, which shows that accidents during lane change belong to a high-frequency type in traffic accidents.

[0003] With the rapid development of intelligent transportation technology, the accurate prediction of lane change behavior has become increasingly important in improving road safety and reducing accidents. In the vehicle networking environment, vehicles can share lane change intentions through mutual communication to optimize traffic flow and reduce the risk of vehicle conflicts; in the field of autonomous driving, lane change prediction is a core link to ensure the safe and smooth driving of vehicles. By using an intelligent system to evaluate the surrounding environment in real time and make accurate lane change decisions, the safety and operating efficiency of the autonomous driving system can be significantly improved.

[0004] The lane change behavior is affected by many factors, and the driving style is one of the key factors. The driving style reflects the driver's operation habits, reaction speed, and adaptability to the traffic environment. Currently, it can be identified through methods such as questionnaires and objective investigations based on vehicle driving characteristic values. The latter covers various ways such as using on-vehicle instruments, mobile phones, driving simulators, and vehicle trajectory data.

[0005] In existing lane change prediction models, most models only focus on the style of the lane-changing vehicle itself and complete the classification of driving styles through factors such as position, speed, acceleration, statistical value information, vehicle type, steering angle, and heading angle. When studying the influence of surrounding vehicles on lane change behavior, the input variables of most models mainly focus on relative features such as position difference, speed difference, and acceleration difference, but ignore the important influence of the driving styles of surrounding vehicles on the lane change process. In actual driving scenarios, drivers with different driving styles will have completely different behavioral performances when facing the lane change of other vehicles. Aggressive drivers may accelerate to obstruct, while conservative drivers will take the initiative to give way, which fully shows that the driving styles of surrounding vehicles will significantly affect the decision-making of lane-changing vehicles. Existing models do not fully consider this key factor of the driving styles of surrounding vehicles, making it difficult to comprehensively understand lane change behavior and resulting in limited prediction accuracy. Summary of the Invention

[0006] Based on this, it is necessary to provide a lane change behavior prediction method, system, device, and medium for the above technical problems.

[0007] An embodiment of the present invention provides a lane change behavior prediction method, including: At different window times starting from the starting time and increasing by time intervals, respectively extract the traffic flow density on the road, the proportion of trucks on the road, and the vehicle length and vehicle feature information of the lane-changing vehicle with a single lane change behavior and surrounding vehicles from the HighD vehicle trajectory dataset; Use Spearman correlation to analyze the vehicle feature information to obtain statistical values, which are used to describe the distribution and fluctuation of vehicle feature information from different perspectives; Take the vehicle length of the lane-changing vehicle and surrounding vehicles, the traffic flow density on the road, the proportion of trucks on the road, and the statistical values as a dataset, and perform clustering analysis on the dataset through the K-means clustering algorithm to divide the driving styles of the lane-changing vehicle and surrounding vehicles into conservative, general, and aggressive types; Adopt the one-way analysis of variance method to calculate the probability value P of each data in the dataset after driving style classification, and take the data with the probability value P less than the set threshold as the key input variables; use the two-layer convolutional neural network 2-CNN as the lane change prediction model, and train the lane change prediction model through the key input variables to obtain the trained lane change prediction model; Respectively input the key input variables at different window times into the trained lane change prediction model for prediction, calculate the F1 value according to the prediction results corresponding to each window time, and take the window time with the maximum F1 value as the optimal window time; During the lane change prediction process, extract the image data of the vehicle to be lane changed and surrounding vehicles at the optimal window time, input the extracted data into the trained lane change prediction model to obtain the lane change behavior prediction result, and perform lane change according to the lane change behavior prediction result.

[0008] Optionally, the surrounding vehicles include: the vehicle in front of the lane where the lane-changing vehicle is located, the vehicle behind in the target lane, and the vehicle in front in the target lane. The vehicle feature information includes the positions, speeds, and accelerations of the lane-changing vehicle and surrounding vehicles, which are divided into vehicle own feature values and associated feature values, specifically including: The vehicle own feature values include: the position x of the vehicle in the x direction, the position y of the vehicle in the y direction, the speed xVelocity of the vehicle in the x direction, the speed yVelocity of the vehicle in the y direction, the acceleration xAcceleration of the vehicle in the x direction, the acceleration in the y direction, and the yAcceleration of the vehicle in the y direction; The associated eigenvalue includes: the distance difference xDiffPreceding in the x-direction between the lane-changing vehicle and the vehicle in front, the distance difference yDiffPreceding in the y-direction between the lane-changing vehicle and the vehicle in front, the speed difference xVelDiffPreceding in the x-direction between the lane-changing vehicle and the vehicle in front, the speed difference yVelDiffPreceding in the y-direction between the lane-changing vehicle and the vehicle in front, the acceleration difference xAccDiffPreceding in the x-direction between the lane-changing vehicle and the vehicle in front, the acceleration difference yAccDiffPreceding in the y-direction between the lane-changing vehicle and the vehicle in front, the distance difference xDiffSidePreceding in the x-direction between the lane-changing vehicle and the vehicle in front in the target lane, the distance difference yDiffSidePreceding in the y-direction between the lane-changing vehicle and the vehicle in front in the target lane, the speed difference xVelDiffSidePreceding in the x-direction between the lane-changing vehicle and the vehicle in front in the target lane, the speed difference yVelDiffSidePreceding in the y-direction between the lane-changing vehicle and the vehicle in front in the target lane, the acceleration difference xAccDiffSidePreceding in the x-direction between the lane-changing vehicle and the vehicle in front in the target lane, the acceleration difference yAccDiffSidePreceding in the y-direction between the lane-changing vehicle and the vehicle in front in the target lane, the distance difference xDiffSideFollowing in the x-direction between the lane-changing vehicle and the vehicle behind in the target lane, the distance difference yDiffSideFollowing in the y-direction between the lane-changing vehicle and the vehicle behind in the target lane, the speed difference xVelDiffSideFollowing in the x-direction between the lane-changing vehicle and the vehicle behind in the target lane, the speed difference yVelDiffSideFollowing in the y-direction between the lane-changing vehicle and the vehicle behind in the target lane, the acceleration difference xAccDiffSideFollowing in the x-direction between the lane-changing vehicle and the vehicle behind in the target lane, and the acceleration difference yAccDiffSideFollowing in the y-direction between the lane-changing vehicle and the vehicle behind in the target lane.

[0009] Optionally, the Spearman correlation is used to analyze the vehicle feature information to obtain statistical values, specifically including: Using Spearman correlation analysis to screen out multiple eigenvalues from the vehicle feature information; Calculating the statistical values of multiple eigenvalues, where the statistical values include the mean, standard deviation, coefficient of variation, mean absolute deviation, and interquartile coefficient of variation; Among them, the mean is used to represent the central position of the eigenvalue, and the calculation formula is: ; The standard deviation S devUsed to measure the dispersion degree of eigenvalues, and the calculation formula is: ; Coefficient of variation C v , used to represent the ratio of the standard deviation to the mean value, reflecting the relative volatility of eigenvalues, and the calculation formula is: ; Mean absolute deviation D mean , and the calculation formula is: ; Interquartile coefficient of variation Q cv Used to characterize the variability of multiple eigenvalues, and the calculation formula is: ; Among them, x i is the i th eigenvalue, n is the total number of eigenvalues, is the average value of all eigenvalues, S dev is the standard deviation, SD is the standard deviation, Q 1 is the first quartile, Q 2 is the second quartile, Q 3 is the third quartile.

[0010] Optionally, use the double-layer convolutional neural network 2-CNN as the lane-changing prediction model, which specifically includes: Determine the search range of the hyperparameters of the double-layer convolutional neural network 2-CNN; the hyperparameters of the double-layer convolutional neural network 2-CNN include: the number of convolutional kernels, the size of the convolutional kernels, the number of neurons in the fully connected layer, and the learning rate; Use the Optuna method for hyperparameter search to determine the values of each hyperparameter of the double-layer convolutional neural network 2-CNN, forming different hyperparameter combinations; Perform cross-validation on each hyperparameter combination to maximize the accuracy of the double-layer convolutional neural network 2-CNN, and obtain the optimal hyperparameter combination; Use the double-layer convolutional neural network 2-CNN corresponding to the optimal hyperparameter combination as the double-layer convolutional neural network 2-CNN after hyperparameter optimization, and use the double-layer convolutional neural network 2-CNN after hyperparameter optimization as the lane-changing prediction model.

[0011] Optionally, train the lane-changing prediction model through key input variables to obtain the trained lane-changing prediction model, which specifically includes: The vehicles that actually change lanes and the vehicles that do not actually change lanes extracted from the HighD dataset are used as actual labels; The lane change prediction model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, and a fully connected layer; The key input variables are received through the input layer; The local features of the key input variables are extracted through the first convolutional layer and the second convolutional layer; The local features are downsampled through the pooling layer to reduce the feature dimension; The downsampled local features are classified through the fully connected layer to obtain the lane change behavior prediction result; The lane change behavior prediction result is compared with the actual label, and the loss function is calculated to measure the accuracy of the model prediction; Through the backpropagation algorithm, the gradient of the loss function with respect to the model parameters is calculated; The model parameters are updated according to the calculated gradient using the optimization algorithm, and iterative training is performed to obtain the trained lane change prediction model.

[0012] Optionally, the image data of the vehicle to be lane-changed and the surrounding vehicles are extracted under the optimal window time, and the extracted data are input into the trained lane change prediction model to obtain the lane change behavior prediction result, which specifically includes: The data extracted from the image data of the vehicle to be lane-changed and the surrounding vehicles are received through the input layer; The primary features of the data are extracted through the first convolutional layer, and the formula is expressed as: ; Among them, is the convolutional kernel, is the input image, * represents the convolution operation, is the bias term, is the activation function; The ReLU activation function is used through the activation layer to increase the nonlinearity of the primary features, and the formula is expressed as: ; The primary features are further feature-extracted through the second convolutional layer to obtain the final features, and the formula is expressed as: ; The final features are downsampled through the pooling layer to reduce the feature dimension of the final features, and the formula is expressed as: ; Among them, A represents the number of layers, F represents the feature values of each layer; The downsampled final features are classified through the fully connected layer to obtain the lane change behavior prediction result.

[0013] An embodiment of the present invention further provides a lane-changing behavior prediction system, including: An information acquisition module, configured to extract the traffic flow density on the road, the proportion of trucks on the road, and the vehicle lengths and vehicle feature information of the lane-changing vehicle with a lane-changing behavior and surrounding vehicles from the HighD vehicle trajectory dataset at different window times starting from the start time and incrementing by time intervals; A data screening module, configured to analyze the vehicle feature information using Spearman correlation to obtain statistical values, which are used to describe the distribution and fluctuation of the vehicle feature information from different perspectives; A clustering analysis module, configured to use the vehicle lengths of the lane-changing vehicle and surrounding vehicles, the traffic flow density on the road, the proportion of trucks on the road, and the statistical values as a dataset, and perform clustering analysis on the dataset through the K-means clustering algorithm to divide the driving styles of the lane-changing vehicle and surrounding vehicles into conservative, general, and aggressive types; A model training module, configured to calculate the probability value P of each data in the dataset after the driving style division using the one-way analysis of variance method, and use the data with the probability value P less than the set threshold as the key input variables; use the two-layer convolutional neural network 2-CNN as the lane-changing prediction model, and train the lane-changing prediction model through the key input variables to obtain the trained lane-changing prediction model; A window time screening module, configured to input the key input variables at different window times into the trained lane-changing prediction model for prediction respectively, calculate the F1 value according to the prediction results corresponding to each window time, and use the window time with the maximum F1 value as the optimal window time; A prediction module, configured to, during the lane-changing prediction process, extract the image data of the vehicle to be lane-changed and surrounding vehicles at the optimal window time, input the extracted data into the trained lane-changing prediction model to obtain the lane-changing behavior prediction result, and perform lane-changing according to the lane-changing behavior prediction result.

[0014] An embodiment of the present invention further provides a computer device, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps of the above-mentioned lane-changing behavior prediction method are implemented.

[0015] An embodiment of the present invention further provides a storage medium, on which a computer program is stored. The computer program is characterized in that when it is executed by a processor, the steps of the above-mentioned lane-changing behavior prediction method are implemented.

[0016] The above-mentioned lane-changing behavior prediction method, system, device, and medium provided by the embodiments of the present invention have the following beneficial effects compared with the prior art: The present invention enriches the input variables of the lane - changing prediction model by quantifying the driving styles of surrounding vehicles and introducing them, along with vehicle length, traffic flow information, etc., as new inputs; classifies the driving styles of lane - changing vehicles and surrounding vehicles using the principal component analysis method and the K - means clustering method to more accurately depict driving behavior characteristics; and, based on driving style characteristics, uses a double - layer convolutional neural network model to make predictions at different window times, exploring the optimal prediction window time and model, which can significantly improve the accuracy of lane - changing behavior prediction and provide strong support for the optimization and application of intelligent transportation technologies Description of the Drawings

[0017] Figure 1 It is a schematic flowchart of a lane - changing behavior prediction method provided in an embodiment; Figure 2 It is a schematic diagram for determining the lane - changing starting point of a lane - changing behavior prediction method provided in an embodiment; Figure 3 It is a shooting scenario diagram of HighD data of a lane - changing behavior prediction method provided in an embodiment; Figure 4 It is a diagram of the positional relationship between the host vehicle and surrounding vehicles of a lane - changing behavior prediction method provided in an embodiment; Figure 5 It is a heat map of the Spearman correlation analysis of 24 feature variables of a lane - changing behavior prediction method provided in an embodiment; Figure 6 It is a diagram of the clustering results of driving styles at different window times of a lane - changing behavior prediction method provided in an embodiment; Figure 7 It is a diagram of the F - value and P - value of the ANOVA test for feature variables at different window times of a lane - changing behavior prediction method provided in an embodiment; Figure 8 It is a diagram of the index trend of the LSTM and 2 - CNN models without considering driving style in a lane - changing behavior prediction method provided in an embodiment; Figure 9 It is a diagram of the index trend of the LSTM and 2 - CNN models considering driving style in a lane - changing behavior prediction method provided in an embodiment; Figure 10 It is a diagram of the performance improvement of the F1 - value of the LSTM and 2 - CNN models before and after considering driving style in a lane - changing behavior prediction method provided in an embodiment; Figure 11 It is a diagram of the relationship between SHAP values and single feature variables at different window times of a lane - changing behavior prediction method provided in an embodiment. Detailed Embodiments

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0019] In the field of free lane - changing prediction on highways, the existing free lane - changing prediction models have significant defects. On the one hand, when analyzing lane - changing behavior, most models overly focus on the driving style of the lane - changing vehicle itself. When considering the influence of surrounding vehicles, they only focus on relative features such as position difference, speed difference, and acceleration difference, seriously neglecting the key factor of the driving style of surrounding vehicles. The driving style of surrounding vehicles can greatly affect their own behavior, and thus influence the decision - making of the lane - changing vehicle. For example, aggressive drivers may impede lane - changing, while conservative drivers will actively cooperate, which greatly reduces the understanding and prediction accuracy of the existing models for lane - changing behavior. On the other hand, in lane - changing prediction research, the selection of the window time has a significant impact on the model performance. Both too short and too long window times will reduce the prediction effect. Although some studies have proposed that a window length of 1 - 5 seconds has certain advantages, there is still a lack of in - depth research on the specific impact of different window times on the lane - changing prediction model considering the driving style of surrounding vehicles. At the same time, the model input variables are not perfect enough, not fully covering important factors such as the type of surrounding vehicles, road traffic flow information, vehicle length, and driving style, unable to comprehensively reflect the actual situation of lane - changing, and reducing the prediction ability of the model.

[0020] In response to the above problems, the present invention provides a brand - new solution. By quantifying the driving style of surrounding vehicles and introducing it, together with vehicle length, traffic flow information, etc., as new inputs into the lane - changing prediction model, the input variables of the model are enriched. The principal component analysis method and the K - means clustering method are used to classify the driving styles of the lane - changing vehicle and surrounding vehicles, more accurately depicting the driving behavior characteristics. At the same time, under different window times, based on the driving style characteristics, the long short - term memory network and the double - layer convolutional neural network model are used for prediction to explore the optimal prediction window time and model. In addition, through SHAP value analysis of the correlation between each input feature and lane - changing prediction, the influence of the driving style of surrounding vehicles, traffic flow information, and different window times on lane - changing behavior prediction is deeply explored, significantly improving the accuracy of lane - changing behavior prediction and providing strong support for the optimization and application of intelligent transportation technology.

[0021] In one embodiment, a lane - changing behavior prediction method is provided, as Figure 1 shown, and the method includes: Extract the traffic flow density on the road, the proportion of trucks on the road, and the vehicle length and vehicle feature information of the lane-changing vehicle with a single lane-changing behavior and surrounding vehicles from the HighD vehicle trajectory dataset at different window times starting from the start time and incrementing by time intervals.

[0022] Analyze the vehicle feature information using Spearman correlation to obtain statistical values, which are used to describe the distribution and fluctuation of vehicle feature information from different perspectives.

[0023] Use the vehicle length of the lane-changing vehicle and surrounding vehicles, the traffic flow density on the road, the proportion of trucks on the road, and the statistical values as a dataset, and perform clustering analysis on the dataset through the K-means clustering algorithm to divide the driving styles of the lane-changing vehicle and surrounding vehicles into conservative, general, and aggressive types.

[0024] Use the one-way ANOVA method to calculate the probability value P of each data in the dataset after the driving style classification, and use the data with the probability value P less than the set threshold as the key input variables; use the two-layer convolutional neural network 2-CNN as the lane-changing prediction model, and train the lane-changing prediction model through the key input variables to obtain the trained lane-changing prediction model.

[0025] Input the key input variables at different window times into the trained lane-changing prediction model for prediction respectively, calculate the F1 value according to the prediction results corresponding to each window time, and use the window time with the maximum F1 value as the optimal window time.

[0026] During the lane-changing prediction process, extract the image data of the vehicle to be lane-changed and surrounding vehicles at the optimal window time, input the extracted data into the trained lane-changing prediction model to obtain the lane-changing behavior prediction result, and perform lane-changing according to the lane-changing behavior prediction result.

[0027] The specific implementation is as follows: 1. Data acquisition and preprocessing This invention uses the HighD vehicle trajectory dataset, as Figure 3 shown. This dataset collects vehicle trajectories on German highways by drones under specific conditions, contains rich information such as vehicle category, driving direction, speed, acceleration, lane position, etc., and covers a large number of lane-changing behavior records. As Figure 2 and Figure 4 shown, this study selects the video data at specific locations and numbers, focuses on the vehicles with a single lane-changing behavior and the vehicles that do not change lanes throughout the journey, and includes the surrounding vehicles of the vehicles with a single lane-changing behavior in the research scope.

[0028] Table 1 Symbol and description of the relationship feature variables between the host vehicle and its surrounding vehicles For each frame of the above 24 eigenvalue data intercepted from the HighD dataset at different window times starting from 1 second and increasing by 0.5 seconds each time, it covers the position, speed, and acceleration of the vehicle in the x and y directions, as well as the distance difference, speed difference, and acceleration difference from surrounding vehicles. For these eigenvalues, five statistical values, namely the mean, standard deviation, coefficient of variation, mean absolute deviation, and interquartile coefficient of variation, are calculated to describe the distribution and fluctuation of the eigenvalues from different perspectives.

[0029] Table 2 Statistical value characteristic variable symbols and descriptions In Table 2: x i : The i th eigenvalue; n : The total number of eigenvalues; : The mean of all eigenvalues, reflecting the central position of the data; S dev : Standard deviation, used to measure the degree of data dispersion; SD : Standard deviation; Q 1: The first quartile (also known as the lower quartile), such that 25% of the data falls below it; Q 2: The second quartile (i.e., the median), such that 50% of the data falls below it; Q 3: The third quartile (also known as the upper quartile), such that 75% of the data falls below it.

[0030] As Figure 5 shown, using Spearman correlation analysis, 21 highly representative eigenvalues are selected, and redundant characteristic variables with a correlation greater than 0.8 are removed. The meaning of highly representative eigenvalues is: characteristic variables with a correlation less than 0.8, that is, if the correlation between two eigenvalues is greater than 0.8, it means the data of these two eigenvalues is similar. Therefore, only one of them needs to be selected, and the other is excluded. So, finally, 21 eigenvalues are obtained.

[0031] Then, five statistics of these 21 eigenvalues are calculated to obtain 105 new characteristic variables. Further, the vehicle length information and corresponding traffic flow information of lane-changing and non-lane-changing vehicles and the three surrounding vehicles are extracted to form a dataset containing 111 characteristic variables.

[0032] Table 3 Vehicle length characteristic variable symbols and descriptions 4.2. Driving Style Analysis and Feature Variable Determination The sorted dataset is normalized so that all features are in the same dimension. The principal component analysis method is used for data dimensionality reduction to simplify the data structure, reduce computational complexity and storage requirements, and at the same time remove redundancy and noise. As Figure 6 shown, the K-means clustering algorithm is used to perform clustering analysis on the dimensionality-reduced data, and the driving styles are divided into three categories: conservative, general, and aggressive, which are represented by 0, 1, and 2 respectively.

[0033] As Figure 7 shown, the one-way analysis of variance (ANOVA) method is adopted to calculate the F value and the probability value P. At the significance level of 0.05, the feature variables that have a significant impact on the lane-changing behavior prediction are screened out (the feature variables with a probability value P < 0.05 have a significant impact on the lane-changing behavior prediction). The number of finally selected feature variables is different under different window times. For example, 64 feature variables are selected for the 1-second window time, providing key input variables for the subsequent prediction model.

[0034] 4.3. Prediction Model Construction and Training The present invention selects the long short-term memory network (LSTM) and the two-layer convolutional neural network (2-CNN) The long short-term memory network LSTM is a special recurrent neural network (RNN) that can learn long-term dependence information. Inside the LSTM, there are three gate structures: the forget gate, the input gate, and the output gate, as well as a cell state. These structures enable the LSTM to perform better than the traditional RNN when processing data with long-term dependence relationships. When using the LSTM for lane-changing behavior prediction, first, data preprocessing is required, and then feature extraction is performed to extract features from the original data that are helpful for predicting the lane-changing behavior. Finally, the data is normalized to accelerate the model training speed.

[0035] When constructing the LSTM model, the input layer is responsible for receiving time series data with a shape of [number of samples, number of time steps, number of features]. Then, through the forget gate, the input gate, the cell state update, and the output gate, the transmission and processing of information are realized. The specific formulas are as follows:

[0036] ; The two - layer convolutional neural network (2 - CNN, Double Convolutional Convolutional Neural Network) is a specially designed convolutional neural network (CNN) architecture. It extends the standard CNN by introducing two consecutive convolutional layers to enhance the network's feature extraction ability. It has strong local feature extraction ability and can effectively capture local features in the data. In lane - changing behavior prediction, the spatial relationships of vehicles (such as the distance from other vehicles, relative positions, etc.) are key factors. The architecture of the double convolutional layer can further enhance this feature and capture more detailed local information. The following is the basic architecture of the double convolutional CNN:

[0037] 1) Input layer: Receives data extracted from the image data of the vehicle to change lanes and surrounding vehicles.

[0038] 2) First convolutional layer: Uses a set of convolutional kernels to extract primary features of the data.

[0039] ; Among them, is the convolutional kernel, is the input image, * represents the convolution operation, is the bias term, is the activation function.

[0040] 3) Activation layer: Applies the ReLU activation function to increase non - linearity.

[0041] ; 4) Second convolutional layer: Further extracts features based on the first convolutional layer to obtain the final features.

[0042] ; 5) Pooling layer: Performs downsampling on the final features to reduce the feature dimension of the final features while retaining important information.

[0043] ; 6) Fully - connected layer: Classifies the final features after downsampling to obtain the lane - changing behavior prediction result.

[0044] ; The core of the dual-convolution CNN lies in its "dual-convolution" structure, that is, after the traditional convolutional layer, instead of directly following the pooling layer, another convolutional layer is added. The design purpose of the dual-convolution CNN is to achieve deeper feature extraction and feature fusion: First, the first convolutional layer is responsible for extracting the feature variable data of the lane-changing vehicle, and these features capture the behavior patterns and state changes of the lane-changing vehicle; Subsequently, the second convolutional layer further operates on this basis to extract the features of surrounding vehicles and capture the interaction information such as the relative position and speed with the lane-changing vehicle. This double-layer structure not only deepens the network's understanding of the lane-changing situation, but also enables the network to learn a richer and more detailed relationship between vehicle behavior and the environment by fusing multi-level features of the lane-changing vehicle and surrounding vehicles, thus showing higher accuracy and stronger generalization ability in the lane-changing behavior prediction task.

[0045] Build a lane-changing prediction model. LSTM can learn long-term dependence information, effectively capture long-term dependence relationships through "memory units" and "gate mechanisms", avoid the problem of gradient disappearance, and has obvious advantages in processing time series data. 2-CNN is extended based on the standard CNN, and enhances the feature extraction ability through two consecutive convolutional layers, and is especially good at capturing local features such as the spatial relationship of vehicles.

[0046] For the LSTM model, a random search cross-validation method is used for hyperparameter optimization. Randomly select a set of parameters from the predefined parameter space for training and evaluation to find the optimal hyperparameter combination. The 2-CNN model uses the Optuna method for hyperparameter search. Optuna is based on the Bayesian optimization algorithm, can more intelligently select the next set of hyperparameter combinations, adjusts the search strategy after each evaluation, and converges to better parameters faster.

[0047] The process of using the Optuna method for hyperparameter search in the 2-CNN model includes: Determine the search range of the hyperparameters of the double-layer convolutional neural network 2-CNN. The hyperparameters of the double-layer convolutional neural network 2-CNN include: the number of convolutional kernels, the size of the convolutional kernels, the number of neurons in the fully connected layer, and the learning rate. Use the Optuna method for hyperparameter search, determine the values of each hyperparameter of the double-layer convolutional neural network 2-CNN, and form different hyperparameter combinations. Perform cross-validation on each hyperparameter combination to maximize the accuracy of the double-layer convolutional neural network 2-CNN and obtain the optimal hyperparameter combination. Use the double-layer convolutional neural network 2-CNN corresponding to the optimal hyperparameter combination as the double-layer convolutional neural network 2-CNN after hyperparameter optimization, and use the double-layer convolutional neural network 2-CNN after hyperparameter optimization as the lane-changing prediction model.

[0048] Table 4 Optimal hyperparameter settings of the LSTM model considering driving style Table 5 Optimal Hyperparameter Settings of the 2-CNN Model Considering Driving Styles 4.4 Lane Change Behavior Prediction and Model Evaluation At different window times, the corresponding feature variables within the window time are extracted from the data of lane-changing vehicles and non-lane-changing vehicles respectively, and input into the trained LSTM and 2-CNN models for lane change behavior prediction.

[0049] Among them, (1) The training process of the LSTM model: 1. Data preparation: Preprocess the input sequence data, usually including normalization and batching, to meet the input requirements of the model.

[0050] 2. Forward propagation: The input data is processed through each time step of the LSTM network. At each time step, the LSTM cell updates its internal state based on the current input and the hidden state of the previous time step, and generates an output.

[0051] 3. Loss calculation: Compare the output of the LSTM with the actual label, and calculate the loss function to measure the accuracy of the model prediction.

[0052] 4. Backward propagation: Calculate the gradient of the loss function with respect to the model parameters through the backward propagation algorithm. Due to the special structure of the LSTM, the backward propagation needs to be processed by unfolding over time for time series data.

[0053] 5. Parameter update: Use the optimization algorithm to update the model parameters according to the calculated gradient to reduce the value of the loss function.

[0054] 6. Iterative training: Repeat the above steps.

[0055] (2) The training process of the 2-CNN model: 1. Data preparation: Preprocess the input data, including normalization and batching, to meet the input requirements of the model.

[0056] 2. Forward propagation: The input data is processed through the convolutional layer, pooling layer, and fully connected layer. The convolutional layer extracts local features, the pooling layer performs downsampling to reduce the feature dimension, and the fully connected layer is used for the final classification or regression task.

[0057] 3. Loss calculation: Compare the output of the model with the actual label, and calculate the loss function to measure the accuracy of the model prediction.

[0058] 4. Backward propagation: Calculate the gradient of the loss function with respect to the model parameters through the backward propagation algorithm.

[0059] 5. Parameter Update: Use an optimization algorithm to update the model parameters based on the calculated gradients to reduce the value of the loss function.

[0060] 6. Iterative Training: Repeat the above steps.

[0061] The implementation process includes: Vehicles that actually change lanes and vehicles that do not actually change lanes extracted from the HighD dataset are used as actual labels. The lane change prediction model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, and a fully connected layer. Key input variables are received through the input layer. Local features of the key input variables are extracted through the first convolutional layer and the second convolutional layer. The local features are downsampled through the pooling layer to reduce the feature dimension. The downsampled local features are classified through the fully connected layer to obtain the lane change behavior prediction result. The lane change behavior prediction result is compared with the actual label, and the loss function is calculated to measure the accuracy of the model prediction. Through the backpropagation algorithm, the gradient of the loss function with respect to the model parameters is calculated. Use an optimization algorithm to update the model parameters based on the calculated gradients, and perform iterative training to obtain the trained lane change prediction model.

[0062] (3)Prediction Process of Long Short-Term Memory Network (LSTM): 1. Data Preprocessing: Normalize the input time series data to improve the convergence speed and prediction accuracy of the model.

[0063] 2. Input Preparation: Divide the preprocessed data into a training set and a test set. In the prediction phase, after training the model using the training set, use the test set for prediction.

[0064] 3. Model Loading: Load the trained LSTM model, including its weights and structure.

[0065] 4. Prediction Input: Input the input data in the test set into the LSTM model.

[0066] 5. Forward Propagation: The LSTM model processes the input data through its internal gate mechanism and memory units to generate a prediction result.

[0067] 6. Output the result.

[0068] (4)Prediction Process of Double-Layer Convolutional Neural Network (2-CNN): 1. Data Preprocessing: Normalize the input data to improve the convergence speed and prediction accuracy of the model.

[0069] 2. Input Preparation: Divide the preprocessed data into a training set and a test set. In the prediction phase, after training the model using the training set, use the test set for prediction.

[0070] 3. Model Loading: Load the trained 2-CNN model, including its weights and structure.

[0071] 4. Prediction Input: Input the input data in the test set into the 2-CNN model.

[0072] 5. Forward Propagation: The 2-CNN model processes the input data through its convolutional layers, pooling layers, and fully connected layers, extracts features, and generates prediction results.

[0073] 6. Output Results.

[0074] Lane-changing behavior occurs when the outputs of both models are [specific value]. Then calculate their accuracy, precision, recall, and F1 value.

[0075] The lane-changing vehicle sets window times of 1s, 1.5s, 2s, 2.5s, 3s, 3.5s, 4s, 4.5s, and 5s forward from the lane-changing starting point, and the lane-changing starting point is determined by the threshold method. The non-lane-changing vehicle randomly selects a starting point and filters the appropriate window time data according to the rules.

[0076] Use accuracy, precision, recall, and F1 value as evaluation indicators, and calculate these indicators based on the confusion matrix of the binary classification problem. By comparing the performance of the LSTM and 2-CNN models at different window times with and without considering driving style features, evaluate the model performance. At the same time, use SHAP values to analyze the contribution of each feature variable to the prediction accuracy, and intuitively show the importance of features such as the driving style of the vehicle itself, the driving styles of the vehicles in front and behind the target lane, vehicle length, and traffic flow information in the model prediction.

[0077] 4.5. Results Analysis and Application 4.5.1. Results Analysis without Considering Driving Style Features As Figure 8 shown, without considering driving style, by comparing the performance of the LSTM and 2-CNN models at different window times, the results show that 2-CNN is slightly better than LSTM at most window times, especially more prominent at short window times.

[0078] Table 6 Prediction Results of LSTM and 2-CNN Models without Considering Driving Style 2-CNN achieved high accuracy, recall, and F1 values ​​at 1 second and 1.5 second window times, which reflect the stability of the model and its ability to accurately predict lane-changing behavior. In particular, at a 1 second window time, the accuracy and F1 value of 2-CNN reached 0.9950 and 0.9809, respectively, which are better than the 0.9821 and 0.9713 of the LSTM model. This shows that 2-CNN has more advantages in short window times and has strong adaptability in predicting lane-changing behavior.

[0079] 4.5.2. Analysis of results considering driving style characteristics like Figure 9 As shown in the figure, under the condition of considering driving style, the LSTM and 2-CNN models have higher accuracy, precision, recall and F1 value at different window times (the F1 value performance improvement is shown in the figure below). Figure 10 As shown in Figure 2). From the perspective of model performance differences under different window times:

[0080] (1) 2-CNN has advantages in short window time: In shorter window time (1 second and 2 seconds), 2-CNN models usually show higher precision, accuracy and F1 value. This is because 2-CNN is more sensitive to feature capture in a short time and can quickly identify changes in driving behavior in a short time, thereby providing accurate predictions. In this case, 2-CNN can capture the details of driving style, so it is often more suitable to choose 2-CNN model for prediction in short window time.

[0081] (2) As the window time increases, LSTM gradually outperforms 2-CNN: When the window time is extended to 3 seconds or more, the recall rate and F1 value of LSTM often exceed those of 2-CNN. In the long window time, driving style and behavior may be continuous and correlated, and the advantage of LSTM for sequence data is revealed here, which can better integrate and understand the changes in driving style over a long period of time. Therefore, in the long window time, LSTM can more accurately capture the continuous change trend of driving behavior.

[0082] (3) 2-CNN is more stable under long windows: In long windows of 4 seconds or more, 2-CNN performs more stably and can maintain a high recall rate and F1 value. This shows that it has strong robustness when dealing with long time series data and is suitable for capturing and predicting continuous changes in driving style. This stability is particularly suitable for driving scenarios with a long duration.

[0083] For different window times of different lengths, the reason for the obvious difference in prediction accuracy may be the difference in the amount of data caused by the length of data collection time. Within a short window of 1 second, the model can focus on high-value information related to lane-changing behavior with less noise, so the prediction accuracy is relatively high. A shorter window can better capture the driver's immediate reactions. Especially when considering driving styles, subtle reaction data helps to accurately predict the next behavior. In a window time of 4.5 seconds and longer, the model accumulates more behavioral information, especially some slowly changing driving characteristics, such as lane-changing preparation or long-term speed adjustment. These information helps to identify chronic driving behavior patterns. Although the increase in window length adds more information, due to the influence of noise, the accuracy may not necessarily exceed that of the short window time.

[0084] When evaluating the effectiveness of the lane-changing intention recognition model, accuracy, precision, recall, and F1-score are selected as evaluation metrics. These metrics are calculated based on the confusion matrix of binary classification problems, and their specific definitions are shown in Table 7.

[0085] Table 7 Confusion Matrix Classification Among them, TP represents the vehicles that change lanes in the dataset, and the prediction result of the lane-changing prediction model is also a lane change; FN represents the vehicles that change lanes in the dataset, but the prediction result of the lane-changing prediction model is no lane change; FP represents the vehicles that do not change lanes in the dataset, but the prediction result of the lane-changing prediction model is a lane change; TN represents the vehicles that do not change lanes in the dataset, and the prediction result of the lane-changing prediction model is also no lane change.

[0086] Precision, also known as the Positive Predictive Value (PPV), measures the proportion of samples that are actually positive among the samples predicted as positive by the model. The formula for precision is as follows:

[0087] ; Recall, also known as the True Positive Rate (TPR) or Sensitivity, measures the proportion of positive samples correctly identified by the model among all actual positive samples. The formula for recall is as follows:

[0088] ; Accuracy is the most intuitive metric to measure the overall performance of the model. It represents the proportion of samples correctly classified by the model among all samples. The formula for accuracy is as follows:

[0089] ; The F1 value is the harmonic mean of precision and recall. It takes both precision and recall into account and is a comprehensive indicator for measuring model performance. The calculation formula for the F1 value is as follows:

[0090] ; Table 8 Prediction Results of LSTM and 2-CNN Models Considering Driving Styles 4.5.3. Model Performance Improvement The F1 value can more comprehensively reflect the performance of the model. From the final F1 value results and performance improvements of each model, for the LSTM model considering driving styles, its prediction accuracy reaches the highest at a window time of 1 second, which is 99.40%. The performance improvement of the model is the most obvious in the window times of 3 seconds and 3.5 seconds, which are 5.20% and 5.73% respectively. This indicates that the LSTM model has the best prediction ability at a relatively short window time and a relatively high model performance improvement at a medium window time. For the 2-CNN model, the accuracy after considering driving styles also reaches the maximum at a window time of 1 second, which is 99.57%. The performance improvement of the model is the most obvious in the window times of 4.5 seconds and 5 seconds, which are 4.74% and 4.71% respectively. The 2-CNN model maintains a relatively high F1 value at a short window time. Especially at the 1-second window time, it demonstrates the ability to effectively capture and multi-layer process the information of the vehicle itself and surrounding vehicles, making its performance at a short window time better than that of the LSTM. Through comparison, it can be seen that in the case of considering driving styles, medium and long window times (3 - 5 seconds) are more effective in improving the prediction accuracy of the model.

[0091] The t-test is a statistical method used to compare whether the difference in the means of two groups of data is significant, and is usually used to verify whether two groups of samples come from the same population. In this study, the t-test was used to test whether the difference in F1 values between considering driving styles and not considering driving styles in different models is statistically significant. The results show that the t-statistic value of the LSTM model is -4.1388 and the p-value is 0.0008, while the t-statistic value of the 2-CNN model is -5.4775 and the p-value is 0.0001. The p-values of both are much smaller than 0.05, indicating that whether it is the LSTM or the 2-CNN model, there are significant differences statistically between considering driving styles and not considering driving styles. This shows that both the LSTM and 2-CNN models exhibit significant performance differences between considering driving styles and not considering driving styles, but there are differences in the significance level and the magnitude of performance improvement between the two.

[0092] Although both show performance improvements after the introduction of driving styles, in terms of the absolute value of the t-statistic, the performance improvement of the 2-CNN model is more significant after considering driving styles. When the 2-CNN model considers driving styles, the magnitude of its performance improvement is larger and the absolute value of the t-statistic is smaller, indicating that driving styles have a stronger impact on this model. This may be related to the structural characteristics of the 2-CNN model. When extracting traffic behavior features, it relies more on fine-grained information. As an additional input feature, driving styles can more effectively enhance the model's recognition ability. In contrast, although the LSTM model also shows a significant performance improvement when considering driving styles, compared with the 2-CNN, the magnitude of the performance improvement of the LSTM is smaller. This may be because its own ability to process time series data is stronger, and the existing network structure can better capture the temporal patterns of driving behaviors. Therefore, the marginal effect of the introduction of driving styles on its performance improvement is relatively small.

[0093] There are differences in the prediction accuracy and stability of LSTM and 2-CNN at different window times, which may be closely related to their characteristics of processing time series. In a short window time, LSTM focuses on the driver's real-time reactions, so its performance is relatively stable. However, in a longer window time, the longer the behavioral sequence captured by LSTM, the more noise there is, resulting in fluctuating performance at different window periods. The 2-CNN can make full use of local features in a short window period (such as 1 second) to capture the subtle changes in the driver's instantaneous actions, so the accuracy is relatively high. As the window period gradually increases, the ability of the 2-CNN to recognize short-term fluctuations decreases, and the prediction accuracy continuously declines. Starting from 4.5 seconds, the 2-CNN can capture some more persistent behavioral features, gradually adapt to longer-term behavioral patterns, and rediscover relatively stable features, so the prediction accuracy rises again at this window time.

[0094] Generally speaking, the performance of both the 2-CNN and LSTM shows significant improvements at each window time after considering driving styles, and different window time adaptability characteristics are shown. Short time and long window time are suitable for using the 2-CNN model for real-time response prediction, while the LSTM is more suitable for capturing the continuous changes of driving behaviors at medium window times. This rule provides a reference for the optimal selection of driving behavior prediction models, and the appropriate model and window time can be selected according to the specific scenario requirements to improve the prediction effect.

[0095] 4.5.4. SHAP-based Analysis of Influencing Factors As Figure 11As shown, by summarizing and visualizing the SHAP values of all features within each window period, a feature importance graph can be generated. The graph intuitively shows that the top 20 features have a relatively high influence in the overall model and are arranged in descending order.

[0096] From the analysis results of comparing the SHAP values for each window time, it can be seen that the influence rankings of these feature variables such as 112 (the driving style of the vehicle itself), 115 (the driving style of the vehicle behind in the target lane), 108 (the length of the vehicle in front in the target lane), 111 (the proportion of trucks in the traffic flow), 110 (the density of the traffic flow), 113 (the driving style of the vehicle in front), 107 (the length of the vehicle in front) are relatively high. This indicates that the newly added vehicle driving style, vehicle length, and traffic flow information in this study have a greater impact when the model predicts whether a lane change behavior will occur, and to a certain extent, it helps to improve the accuracy of model prediction.

[0097] In the prior art, a highway lane change decision-making method and system considering the driving style of surrounding vehicles (abbreviated as Method A) is proposed. Its main content is as follows: Classify the driving styles of surrounding vehicles according to different degrees of aggressiveness, and determine the characteristic parameters representing the horizontal and vertical characteristics of different driving styles; Based on reinforcement learning, construct a decision-making model for surrounding vehicles with different driving styles to predict the behaviors of surrounding vehicles with different driving styles. In the first simulation environment, use the first reward function to train the decision-making model for surrounding vehicles with different driving styles. Among them, the first reward function considers the characteristic parameters representing the horizontal and vertical characteristics of different driving styles, which can make the trained decision-making model for surrounding vehicles exhibit different driving styles under different characteristic parameters. Then, based on reinforcement learning, construct a highway lane change decision-making model considering the driving style of surrounding vehicles. In the second simulation environment, use the second reward function to train the highway lane change decision-making model. Among them, the second simulation environment includes surrounding vehicles with different driving styles, and the behavior decisions of surrounding vehicles with different driving styles are obtained through the trained decision-making models for surrounding vehicles with different driving styles, which can simulate complex traffic scenarios. And the second reward function considers the driving style of surrounding vehicles, making the highway lane change decision-making model have better generalization for surrounding vehicles with different driving styles. Finally, use the trained highway lane change decision-making model to make lane change decisions, improving the robustness and safety of lane change decisions.

[0098] In the prior art, a method and system for vehicle trajectory prediction and behavior decision-making considering driving style (abbreviated as Method B) are also proposed. The main content is as follows: First, a driving behavior data set is obtained and processed to obtain a balanced data set for vehicle trajectory prediction. The driving style vector labels obtained by classifying the original driving behavior data set based on a driving style classifier are incorporated into this balanced data set, which provides a richer information dimension for subsequent analysis. Then, a long short-term memory network based on Bayesian optimization and fused with a discrete cosine transform attention mechanism is constructed to perform high-precision vehicle trajectory prediction. During the network construction process, a single-layer convolutional neural network is first used to extract potential features from the input target vehicle state observation feature vector, and the target vehicle state observation feature vector Input(t) is composed of multi-dimensional information. Then, the extracted potential features are input into the long short-term memory network, and a frequency-enhanced channel attention mechanism based on the discrete cosine transform is introduced to enhance the network's ability to capture key features. At the same time, the Bayesian optimization algorithm is used to optimize the hyperparameters of the model, and the mean square error function is defined as the objective function to be minimized by the Bayesian optimization algorithm. In this way, the overall performance of the model is improved, and more accurate vehicle trajectory prediction is achieved. Finally, combining the safety factor, comfort factor, efficiency factor, gain factor in the multi-lane lane-changing scenario, as well as the driving style vector label and vehicle trajectory prediction information, a deep learning model is used to implement lane-changing behavior decision-making.

[0099] The comparison between the present invention and Method A and Method B is as follows: (1) The present invention constructs a system that fully considers the influence of the driving styles of surrounding vehicles. Most previous studies have focused on the lane-changing vehicle itself, while the present invention notices the key influence of the driving styles of surrounding vehicles on lane-changing behavior. By carefully classifying driving styles into conservative, general, and aggressive types and incorporating them into the lane-changing prediction model, a quantitative analysis of the influence mechanism is achieved. This enables the model to highly restore the complex interaction relationships in the real traffic scenario, greatly improving the accuracy and reliability of prediction, providing a more accurate decision-making basis for the autonomous driving system, and effectively reducing the risk of traffic accidents.

[0100] (2) For the lane-changing vehicle, the threshold method is used to accurately determine the lane-changing starting point, and then multiple groups of window times with different durations are innovatively set to comprehensively capture the dynamic changes of the vehicle before lane-changing. At different window times from 1s to 5s, the LSTM and 2-CNN models are combined for prediction, and multi-index comprehensive evaluation is used to successfully determine the optimal window time in various scenarios. Compared with the selection of a single window time in previous studies, this study considers more comprehensively, and the finally selected optimal window time is more persuasive and more helpful for improving the accuracy of model prediction.

[0101] (3) When constructing the model, it is fully considered that the lane-changing behavior of vehicles is closely related to vehicle type characteristics and traffic flow status. Innovatively, key information such as vehicle length, traffic flow density, and the proportion of trucks is incorporated into the feature values. Compared with traditional models that only focus on the basic elements of vehicle trajectories, the present invention greatly enriches the input dimension of the model, enabling the model to more comprehensively and realistically reflect the actual lane-changing situation.

[0102] Based on the same inventive concept, a lane-changing behavior prediction system is proposed, including: An information acquisition module, configured to extract the traffic flow density on the road, the proportion of trucks on the road, and the vehicle lengths and vehicle feature information of the lane-changing vehicles with one lane-changing behavior and surrounding vehicles from the HighD vehicle trajectory dataset at different window times starting from the start time and incrementing by time intervals.

[0103] A data screening module, configured to analyze the vehicle feature information using Spearman correlation to obtain statistical values, which are used to describe the distribution and fluctuation of the vehicle feature information from different perspectives.

[0104] A clustering analysis module, configured to use the vehicle lengths of the lane-changing vehicles and surrounding vehicles, the traffic flow density on the road, the proportion of trucks on the road, and the statistical values as a dataset, and perform clustering analysis on the dataset through the K-means clustering algorithm to divide the driving styles of the lane-changing vehicles and surrounding vehicles into conservative, general, and aggressive types.

[0105] A model training module, configured to calculate the probability value P of each data in the dataset after the driving style division using the one-way analysis of variance method, and use the data with the probability value P less than the set threshold as the key input variables. Taking the double-layer convolutional neural network 2-CNN as the lane-changing prediction model, training the lane-changing prediction model with the key input variables to obtain the trained lane-changing prediction model.

[0106] A window time screening module, configured to input the key input variables at different window times into the trained lane-changing prediction model for prediction respectively, calculate the F1 value according to the prediction results corresponding to each window time, and use the window time with the maximum F1 value as the optimal window time.

[0107] A prediction module, configured to, during the lane-changing prediction process, extract the image data of the vehicle to be lane-changed and surrounding vehicles at the optimal window time, input the extracted data into the trained lane-changing prediction model to obtain the lane-changing behavior prediction result, and perform lane-changing according to the lane-changing behavior prediction result.

[0108] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention.

Claims

1. A lane-changing behavior prediction method, characterized in that, Including: Extract the traffic flow density on the road, the proportion of trucks on the road, and the vehicle length and vehicle feature information of the lane-changing vehicle with a lane-changing behavior and surrounding vehicles from the HighD vehicle trajectory dataset at different window times starting from the start time and increasing at each time interval; Analyze the vehicle feature information using Spearman correlation to obtain statistical values, which are used to describe the distribution and fluctuation of vehicle feature information from different perspectives; Use the vehicle length of the lane-changing vehicle and surrounding vehicles, the traffic flow density on the road, the proportion of trucks on the road, and the statistical values as a dataset, and perform clustering analysis on the dataset through the K-means clustering algorithm to divide the driving styles of the lane-changing vehicle and surrounding vehicles into conservative, general, and aggressive types; Use the one-way ANOVA method to calculate the probability value P of each data in the dataset after the driving style is divided, and use the data with the probability value P less than the set threshold as the key input variables; use the double-layer convolutional neural network 2-CNN as the lane-changing prediction model, and train the lane-changing prediction model with the key input variables to obtain the trained lane-changing prediction model; Input the key input variables at different window times into the trained lane-changing prediction model for prediction respectively, calculate the F1 value according to the prediction results corresponding to each window time, and use the window time with the maximum F1 value as the optimal window time; During the lane-changing prediction process, extract the image data of the vehicle to be lane-changed and surrounding vehicles at the optimal window time, input the extracted data into the trained lane-changing prediction model to obtain the lane-changing behavior prediction result, and perform lane-changing according to the lane-changing behavior prediction result.

2. The lane-changing behavior prediction method according to claim 1, wherein The surrounding vehicles include: the vehicle in front in the lane where the lane-changing vehicle is located, the vehicle behind in the target lane, and the vehicle in front in the target lane. The vehicle feature information includes the positions, speeds, and accelerations of the lane-changing vehicle and surrounding vehicles respectively, and is divided into vehicle feature values and associated feature values, specifically including: The vehicle feature values include: the position x in the x direction of the lane-changing vehicle and surrounding vehicles respectively, the position y in the y direction of the lane-changing vehicle and surrounding vehicles respectively, the speed xVelocity in the x direction of the lane-changing vehicle and surrounding vehicles respectively, the speed yVelocity in the y direction of the lane-changing vehicle and surrounding vehicles respectively, the acceleration xAcceleration in the x direction of the lane-changing vehicle and surrounding vehicles respectively, the lane-changing vehicle and surrounding vehicles respectively, the acceleration in the y direction of the lane-changing vehicle and surrounding vehicles and yAcceleration; The associated eigenvalue includes: the distance difference xDiffPreceding in the x direction between the lane-changing vehicle and the vehicle in front in the same lane, the distance difference yDiffPreceding in the y direction between the lane-changing vehicle and the vehicle in front in the same lane, the speed difference xVelDiffPreceding in the x direction between the lane-changing vehicle and the vehicle in front, the speed difference yVelDiffPreceding in the y direction between the lane-changing vehicle and the vehicle in front in the same lane, the acceleration difference xAccDiffPreceding in the x direction between the lane-changing vehicle and the vehicle in front in the same lane, the acceleration difference yAccDiffPreceding in the y direction between the lane-changing vehicle and the vehicle in front in the same lane, the distance difference xDiffSidePreceding in the x direction between the lane-changing vehicle and the vehicle in front in the target lane, the distance difference yDiffSidePreceding in the y direction between the lane-changing vehicle and the vehicle in front in the target lane, the speed difference xVelDiffSidePreceding in the x direction between the lane-changing vehicle and the vehicle in front in the target lane, the speed difference yVelDiffSidePreceding in the y direction between the lane-changing vehicle and the vehicle in front in the target lane, the acceleration difference xAccDiffSidePreceding in the x direction between the lane-changing vehicle and the vehicle in front in the target lane, the acceleration difference yAccDiffSidePreceding in the y direction between the lane-changing vehicle and the vehicle in front in the target lane, the distance difference xDiffSideFollowing in the x direction between the lane-changing vehicle and the vehicle behind in the target lane, the distance difference yDiffSideFollowing in the y direction between the lane-changing vehicle and the vehicle behind in the target lane, the speed difference xVelDiffSideFollowing in the x direction between the lane-changing vehicle and the vehicle behind in the target lane, the speed difference yVelDiffSideFollowing in the y direction between the lane-changing vehicle and the vehicle behind in the target lane, the acceleration difference xAccDiffSideFollowing in the x direction between the lane-changing vehicle and the vehicle behind in the target lane, and the acceleration difference yAccDiffSideFollowing in the y direction between the lane-changing vehicle and the vehicle behind in the target lane.

3. The lane-changing behavior prediction method according to claim 1, wherein The analysis of vehicle feature information using Spearman correlation to obtain statistical values specifically includes: Using Spearman correlation analysis to screen out multiple eigenvalues from vehicle feature information; Calculating the statistical values of multiple eigenvalues, where the statistical values include mean, standard deviation, coefficient of variation, mean absolute deviation, and interquartile coefficient of variation: wherein, the average value is used to represent the central position of the eigenvalue, and the calculation formula is: ； The standard deviation S dev is used to measure the degree of dispersion of eigenvalue, and the calculation formula is: ； The coefficient of variation C v , which is used to represent the ratio of the standard deviation to the mean value and reflects the relative volatility of the eigenvalue. The calculation formula is as follows: ； The mean absolute deviation D mean , and the calculation formula is as follows: ； The coefficient of variation of quartiles Q cv is used to characterize the variability of multiple eigenvalue, and the calculation formula is: ； wherein, x i is the i th eigenvalue, n is the total number of eigenvalues, is the average value of all eigenvalues, S dev is the standard deviation, SD is the standard deviation, Q 1 is the first quartile, Q 2 is the second quartile, Q 3 is the third quartile.

4. The lane-changing behavior prediction method according to claim 1, wherein Regarding the use of the double-layer convolutional neural network 2-CNN as a lane-changing prediction model, it specifically includes: Determining the search range of the hyperparameters of the double-layer convolutional neural network 2-CNN; the hyperparameters of the double-layer convolutional neural network 2-CNN include: the number of convolutional kernels, the size of the convolutional kernels, the number of neurons in the fully connected layer, and the learning rate; Using the Optuna method for hyperparameter search to determine the values of each hyperparameter of the double-layer convolutional neural network 2-CNN, forming different hyperparameter combinations; Perform cross-validation on each combination of hyperparameters to maximize the accuracy of the two-layer convolutional neural network 2-CNN and obtain the optimal combination of hyperparameters; Use the two-layer convolutional neural network 2-CNN corresponding to the optimal combination of hyperparameters as the two-layer convolutional neural network 2-CNN after hyperparameter optimization, and use the two-layer convolutional neural network 2-CNN after hyperparameter optimization as the lane-changing prediction model.

5. The lane-changing behavior prediction method according to claim 4, characterized in that, Training the lane-changing prediction model through key input variables to obtain the trained lane-changing prediction model, specifically including: Use the vehicles that actually change lanes and the vehicles that do not actually change lanes extracted from the HighD dataset as actual labels; The lane-changing prediction model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, and a fully connected layer; Receive key input variables through the input layer; Extract local features of key input variables through the first convolutional layer and the second convolutional layer; Perform downsampling on the local features through the pooling layer to reduce the feature dimension; Classify the downsampled local features through the fully connected layer to obtain the lane-changing behavior prediction result; Compare the lane-changing behavior prediction result with the actual label and calculate the loss function to measure the accuracy of the model prediction; Calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm; Use the optimization algorithm to update the model parameters according to the calculated gradient, perform iterative training, and obtain the trained lane-changing prediction model.

6. The lane-changing behavior prediction method according to claim 5, characterized in that, Extract the image data of the vehicle to be lane-changed and surrounding vehicles at the optimal window time, and input the extracted data into the trained lane-changing prediction model to obtain the lane-changing behavior prediction result: Receive the data extracted from the image data of the vehicle to be lane-changed and surrounding vehicles through the input layer; Extract the primary features of the data through the first convolutional layer, and the formula is expressed as: ； Among them, is the convolutional kernel, is the input image, * represents the convolution operation, is the bias term, is the activation function; Use the ReLU activation function through the activation layer to increase the non-linearity of the primary features, and the formula is expressed as: ； The primary features are subjected to feature extraction through the second convolutional layer to obtain the final features, and the formula is expressed as: ; Perform downsampling on the final features through the pooling layer to reduce the feature dimension of the final features, and the formula is expressed as: ； Among them, A represents the number of layers, F and represents the eigenvalue of each layer; Classify the final features after downsampling through the fully connected layer to obtain the lane-changing behavior prediction result.

7. A lane-changing behavior prediction system, characterized in that, Including: An information acquisition module for extracting the traffic flow density on the road, the proportion of trucks on the road, and the vehicle lengths and vehicle feature information of the lane-changing vehicle with one lane-changing behavior and surrounding vehicles from the HighD vehicle trajectory dataset at different window times starting from the start time and increasing at each time interval; A data screening module for analyzing the vehicle feature information using Spearman correlation to obtain statistical values, and the statistical values are used to describe the distribution and fluctuation of the vehicle feature information from different perspectives; A clustering analysis module for using the vehicle lengths of the lane-changing vehicle and surrounding vehicles, the traffic flow density on the road, the proportion of trucks on the road, and the statistical values as a dataset, performing clustering analysis on the dataset through the K-means clustering algorithm, and classifying the driving styles of the lane-changing vehicle and surrounding vehicles into conservative, general, and aggressive types; A model training module, which is used to calculate the probability value P of each data in the dataset after driving style classification by using the one-way analysis of variance method, and take the data with the probability value P less than the set threshold as the key input variables; use the double-layer convolutional neural network 2-CNN as the lane-changing prediction model, and train the lane-changing prediction model with the key input variables to obtain the trained lane-changing prediction model; A window time screening module, which is used to input the key input variables under different window times into the trained lane-changing prediction model for prediction respectively, calculate the F1 value according to the prediction results corresponding to each window time, and take the window time with the maximum F1 value as the optimal window time; A prediction module, which is used to extract the image data of the vehicle to be lane-changed and the surrounding vehicles at the optimal window time during the lane-changing prediction process, input the extracted data into the trained lane-changing prediction model to obtain the lane-changing behavior prediction result, and perform lane-changing according to the lane-changing behavior prediction result.

8. A computer device, comprising a memory and a processor, wherein a computer program is stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of a lane-changing behavior prediction method according to any one of claims 1-6.

9. A storage medium, on which a computer program is stored, characterized in that, When the computer program is executed by the processor, it implements the steps of a lane-changing behavior prediction method according to any one of claims 1-6.

Images