Driving style identification method based on lane changing driving data
By introducing Bayesian regularization mechanism into the SOM algorithm, optimizing weight updates and hyperparameter adjustments, the problems of poor clustering effect and poor generalization ability in driving style recognition are solved, and higher driving style recognition accuracy and cluster cohesion are achieved.
Patent Information
- Application Number
- CN202510459908.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-14
- Publication Date
- 2025-05-23
AI Technical Summary
The existing SOM methods have problems such as poor clustering effect, easy overfitting and excessive neuronal contraction in driving style recognition, resulting in poor generalization ability.
Using Bayesian regularization-based self-organized mapping (BR-SOM) algorithm, the clustering effect and generalization ability of SOM are improved by optimizing weight update formulas and Bayesian inference dynamic optimization.
It improves the accuracy of driving style recognition, significantly improves the classification accuracy of driving styles such as radical and conservative, and has a higher clustering and cohesion, which can better process large-scale data.
Smart Images
Figure CN120024337A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of vehicle assisted driving control and relates to a driving style recognition method, and in particular to a driving style recognition method based on lane change driving data. Background Art
[0002] Different drivers have different driving styles, so it is necessary to identify the driver's driving style and adjust the control parameters according to the driver's driving style and vehicle status so that the vehicle's control characteristics meet the driver's personalized needs.
[0003] In order to reduce the difficulty of driving for drivers, reduce the burden of driving for drivers, effectively avoid fatigue driving, improve road traffic safety, and better achieve the purpose of human-machine co-driving, it is necessary to fully understand the physiological characteristics, driving status, driving habits, etc. of drivers, so as to effectively identify driving risks, take effective warnings before accidents occur and take measures to avoid accidents, reduce personal safety injuries and economic losses. At present, autonomous driving is in a stage of continuous improvement and development. Driving style represents the driving behavior of drivers and plays an important role in traffic safety, vehicle energy management and vehicle insurance systems. With the rapid development of intelligent vehicle technology, the relationship between cars, traffic and drivers has become extremely complex. Carrying out driver style recognition and conducting research based on driver behavior understanding will help achieve harmonious interaction and intelligent integration between drivers and intelligent systems, and promote the development of autonomous driving and intelligent transportation industries.
[0004] Different drivers have different driving habits and levels. The current classification of driving styles is mainly divided into rule-based methods, statistical methods and machine learning methods. However, the rules of rule-based methods are too simple to capture complex driving behaviors; statistical methods rely on feature engineering, feature extraction is complex and subjective, it is difficult to process high-dimensional data, and it is easy to lose important information; machine learning methods require a large amount of labeled data, the model is poorly interpretable, and it is difficult to understand the specific content of driving style. The neural network algorithm based on SOM can solve the problems of difficult labeling data and difficult processing of complex driving behavior data in existing driver style recognition methods. However, the traditional SOM method has the following problems: when the number of neurons is large, the same driving behavior samples are easily scattered to different neurons, affecting the clustering effect, and it is easy to overfit the training data, resulting in poor generalization ability. In the absence of additional control parameters, it is inevitable to encounter the problem of excessive shrinkage of neurons. Summary of the invention
[0005] In order to solve the above technical problems, the present invention provides a driving style recognition method based on lane change driving data, which performs driving style recognition based on the SOM algorithm and the BR-DPSOM neural network to solve the problems of poor clustering effect, easy overfitting, excessive neuron shrinkage, etc. of the current SOM method.
[0006] The present invention provides a driving style recognition method based on lane change driving data, comprising the following steps:
[0007] Step 1: Build a driving data collection platform and construct driving scenarios; design lane-changing scenarios under different working conditions;
[0008] Furthermore, in the lane-changing scenario, the expected driving speed of the test vehicle is required to be within the range of 60km / h to 80km / h; there is a certain initial distance between the front vehicle and the test vehicle in the same lane, and the speed of the front vehicle is lower than the expected speed of the test vehicle; at the same time, there is a certain initial distance between the rear vehicle and the test vehicle in the other lane, and the speed of the rear vehicle is close to the expected speed of the test vehicle.
[0009] As a preferred method, three vehicles are set in the lane change scenario, namely the front vehicle, the test vehicle and the rear vehicle. The initial distance between the test vehicle and the front vehicle is 100 meters. The test vehicles are divided into three categories according to their driving speeds. Each category is further divided into six subcategories according to the longitudinal speeds of the front vehicle and the rear vehicle in the lane, for a total of 18 working conditions:
[0010] When the expected speed of the test vehicle is 80km / h, the speed of the leading vehicle is 70km / h or 60km / h, and the speed of the trailing vehicle is 90km / h (100m from the test vehicle), 80km / h (50m from the test vehicle), and 70km / h (20m from the test vehicle).
[0011] When the expected speed of the test vehicle is 70km / h, the speed of the leading vehicle is 60km / h or 50km / h, and the speed of the trailing vehicle is 80km / h (100m from the test vehicle), 70km / h (50m from the test vehicle), and 60km / h (20m from the test vehicle).
[0012] When the expected speed of the test vehicle is 60km / h, the speed of the leading vehicle is 50km / h or 40km / h, and the speed of the rear vehicle is 70km / h (the distance from the test vehicle is 100m), 60km / h (the distance from the test vehicle is 50m), and 50km / h (the distance from the test vehicle is 20m).
[0013] Step 2: Collect lane-changing data of drivers with different driving styles under different working conditions; the lane-changing data includes vehicle status data and vehicle operation data; the vehicle status data includes the longitudinal displacement of the main vehicle (m), the lateral displacement of the main vehicle (m), the longitudinal speed of the main vehicle (m / s), the lateral speed of the main vehicle (m / s), the longitudinal acceleration of the main vehicle (m / s2 ), lateral acceleration of the vehicle (m / s 2 ) one or more of; the vehicle control data includes one or two of the main vehicle steering wheel angle (rad), the main vehicle throttle opening (%);
[0014] Step 3: Collect driving data according to the start and end time of lane change, perform statistical processing on the raw data, and obtain driver style characteristic parameters, including:
[0015] Extracting the lane-changing center time of the main vehicle; the lane-changing center time of the main vehicle is the time when the center of the main vehicle is at the lane-changing point;
[0016] Extract the starting point time of the main vehicle changing lanes;
[0017] Extract the lane-changing end point time of the main vehicle.
[0018] Step 4: Use the principal component analysis method to reduce the dimension of the statistical features of each driving segment to obtain a new principal component after dimension reduction. The principal component is formed by a linear combination of various parameters in the index; then sort the principal components to obtain the principal component contribution rate and cumulative contribution rate, and extract the driving characteristics;
[0019] Furthermore, the principal component analysis method comprises the following steps:
[0020] Step 4.1: Data standardization:
[0021] Suppose the original data set is X, where each row is a sample and each column is a feature; the standardized data is: Among them, μ is the mean vector of each feature, and σ is the standard deviation vector of each feature;
[0022] Step 4.2, calculate the covariance matrix:
[0023] The covariance matrix calculation formula is: Where n is the number of samples, ∑ is a d×d symmetric matrix, d is the number of features, X scaled To standardize the data set;
[0024] Step 4.3, select the main components:
[0025] Select the eigenvectors corresponding to the k largest eigenvalues to form the projection matrix W; W = Q :k ; Among them, Q :k is the first k columns of Q;
[0026] Step 4.4, data projection:
[0027] Project the standardized data into the principal component space, the formula is X pca =X scaledW, where X pca is the data after dimensionality reduction, with a dimension of n×k.
[0028] Step 5: Use the SOM clustering method to cluster each driving feature to form multiple driving style types, and design driving style recognition rules based on cluster centers. The process is as follows:
[0029] Step 5.1, initialize the SOM grid:
[0030] The SOM grid is an m×m two-dimensional grid, and each node j has a weight vector w j , dimension is k; initialize the weight vector w j ;
[0031] Step 5.2, calculate the distance between the input sample and the weight vector:
[0032] For each input sample x i , calculate the weight vector w between it and each node j j The Euclidean distance is calculated as d(x i ,w j )=||x i -w j || 2 ;
[0033] Step 5.3, find the best matching unit (BMU):
[0034] Find the input sample x i The node j with the smallest distance * , the calculation formula is
[0035] Step 5.4. Update BMUj * and the weight vector of its neighboring nodes:
[0036] The formula is Where t is the current iteration number, η(t) is the learning rate, which decays over time; is the neighborhood function, using the Gaussian function, the formula is in, and r j are the best matching unit (BMU) and the position of node j in the grid, respectively, σ(t) is the neighborhood radius, which decays with time;
[0037] Step 5.5, iteration, repeat steps 5.2 to 5.4 until the maximum number of iterations is reached or the weight vector converges.
[0038] Step 6: Analyze the SOM clustering method used in step 5 to improve its limitations; modify and optimize the SOM clustering method based on the Bayesian regularization method; design an improved driving style recognition rule; including the following steps:
[0039] Step 6.1, optimize the SOM weight adjustment formula:
[0040] In the traditional SOM weight update formula, a regularization penalty term is added to define a correction vector. The new correction vector objective function is m k (t) = δE t (t)+∈E w (t), the modified weight adjustment formula, δ and ∈ are hyperparameters, and the hyperparameters δ and ∈ control the weights of the error term and the regularization term respectively; among them, the error term E k (t) is the difference between the input sample and the weight, E k (t) = x k -ω k (t); Regularization term E w (t) is the weighted mean square value, Used to penalize weight complexity; x k represents the kth eigenvalue of the input sample, ω k (t) represents the weight vector of neuron k at time step t, that is, the position of the neuron in the feature space; w ij (t) represents the connection weight from input node i to output node j at time step t, which is a specific element of the weight matrix in the SOM network;
[0041] Step 6.2, Bayesian reasoning dynamic optimization:
[0042] According to Bayesian reasoning, assuming that the hyperparameters δ and ∈ obey Gaussian distribution, the objective function m is solved by maximizing the posterior probability. k (t) The hyperparameters δ and ∈ at their minimum point, i.e. the optimal value of the hyperparameters Where, λ = n-2δ·Trace(T -1 ), T is the Hessian matrix of the modified objective function at the minimum point; the Bayesian method iteratively updates the hyperparameters to gradually make the weight distribution consistent with the probability distribution of the input data;
[0043] Step 6.3, update the SOM neuron position:
[0044] Taking the winning neuron as the benchmark, adjust the positions of other neurons during the iteration process: Among them, δ vk and d vkare the distances in high-dimensional and low-dimensional space respectively, p k (t) is the neuron position, α(t) is the learning rate, p v (t) is the position of the winning neuron.
[0045] Beneficial effects of the present invention:
[0046] The present invention provides a driving style recognition method based on lane-changing driving data, which performs driver style clustering based on Bayesian regularized self-organizing mapping, and can improve the accuracy of driving style recognition: the weight update process is optimized through the Bayesian regularization mechanism, so that similar driving behavior features are more closely clustered in the feature space. In the lane-changing scene data, the classification accuracy of driving styles such as aggressive and conservative is significantly better than that of the traditional SOM algorithm.
[0047] The clustering cohesion of the present invention is higher than that of the traditional SOM algorithm. As the number of grids increases, the cohesion of the two algorithms continues to decrease. Compared with the traditional SOM algorithm, the decreasing trend of the present invention is more gentle. When the number of grids is large enough (or close to the number of samples), most of the samples in the clustering results of the traditional SOM algorithm are scattered on different neurons, and the recognition rate of neurons for samples of the same category is low. In the results of the present invention, neurons can condense several samples of the same category, and the recognition rate of neurons for samples of the same category is high. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 It is a schematic diagram of the overall process of the method of the present invention;
[0049] Figure 2 The clustering result of the embodiment of the present invention is Figure 1 ;
[0050] Figure 3 The clustering result of the embodiment of the present invention is Figure 2 ;
[0051] Figure 4 Comparison of silhouette coefficients between the SOM method and the BR-SOM method of the present invention. DETAILED DESCRIPTION
[0052] like Figure 1 As shown, the present invention provides a driving style recognition method based on lane change driving data, comprising the following steps:
[0053] Step 1: Build a driving data collection platform. In this embodiment, a standard suburban highway scene with two lanes in one direction is constructed through PreScan software.
[0054] Design lane-changing scenarios under different working conditions. In the lane-changing scenarios, the expected driving speed of the test vehicle is required to be within the range of 60km / h to 80km / h. The front vehicle in the same lane has a certain initial distance from the test vehicle, and the speed of the front vehicle is lower than the expected speed of the test vehicle. At the same time, the rear vehicle in the other lane has a certain initial distance from the test vehicle, and the speed of the rear vehicle is close to the expected speed of the test vehicle.
[0055] In this embodiment, three vehicles are set, namely the front vehicle, the test vehicle and the rear vehicle. The initial distance between the test vehicle and the front vehicle is 100 meters. The test vehicles are divided into three categories according to their driving speeds. Each category is further divided into six subcategories according to the longitudinal speeds of the front vehicle and the rear vehicle in the lane. There are 18 working conditions in total:
[0056] When the expected speed of the test vehicle is 80km / h, the speed of the leading vehicle is 70km / h or 60km / h, and the speed of the trailing vehicle is 90km / h (100m from the test vehicle), 80km / h (50m from the test vehicle), and 70km / h (20m from the test vehicle).
[0057] When the expected speed of the test vehicle is 70km / h, the speed of the leading vehicle is 60km / h or 50km / h, and the speed of the trailing vehicle is 80km / h (100m from the test vehicle), 70km / h (50m from the test vehicle), and 60km / h (20m from the test vehicle).
[0058] When the expected speed of the test vehicle is 60km / h, the speed of the leading vehicle is 50km / h or 40km / h, and the speed of the rear vehicle is 70km / h (the distance from the test vehicle is 100m), 60km / h (the distance from the test vehicle is 50m), and 50km / h (the distance from the test vehicle is 20m).
[0059] Step 2: Collect lane-changing data of drivers with different driving styles under different working conditions; the lane-changing data includes vehicle status data and vehicle operation data; the vehicle status data includes the longitudinal displacement of the main vehicle (m), the lateral displacement of the main vehicle (m), the longitudinal speed of the main vehicle (m / s), the lateral speed of the main vehicle (m / s), the longitudinal acceleration of the main vehicle (m / s 2 ), lateral acceleration of the vehicle (m / s 2 ); the vehicle control data includes the steering wheel angle of the main vehicle (rad), the throttle opening of the main vehicle (%);
[0060] Step 3: collecting driving data according to the start and end time of lane change; statistically processing the raw data to obtain driver style characteristic parameters, the steps include:
[0061] Extracting the lane-changing center time of the main vehicle; the lane-changing center time of the main vehicle is the time when the center of the main vehicle is at the lane-changing point;
[0062] Extract the starting point time of the main vehicle changing lanes;
[0063] Extract the lane-changing end point time of the main vehicle.
[0064] Step 4: Use the principal component analysis method to reduce the dimension of the statistical features of each driving segment to obtain the new principal components after dimension reduction. The principal components are formed by the linear combination of the parameters in the index; then sort the principal components to obtain the principal component contribution rate and the cumulative contribution rate, calculate the cumulative variance contribution rate, retain the dimension that explains more than 85% of the variance to obtain the data after dimension reduction, and extract the driving characteristics;
[0065] The principal component analysis method comprises the following steps:
[0066] Step 4.1: Data standardization:
[0067] Suppose the original data set is X, where each row is a sample and each column is a feature; the standardized data is: Among them, μ is the mean vector of each feature, and σ is the standard deviation vector of each feature;
[0068] Step 4.2, calculate the covariance matrix:
[0069] The covariance matrix calculation formula is: Where n is the number of samples, ∑ is a d×d symmetric matrix, d is the number of features, X scaled To standardize the data set;
[0070] Step 4.3, select the main components:
[0071] Select the eigenvectors corresponding to the k largest eigenvalues to form the projection matrix W; W = Q :k ; Among them, Q :k is the first k columns of Q;
[0072] Step 4.4, data projection:
[0073] Project the standardized data into the principal component space, the formula is X pca =X scaled W, where X pca is the data after dimensionality reduction, with a dimension of n×k.
[0074] Step 5: Use the SOM clustering method to cluster each driving feature to form multiple driving style types, including aggressive, stable, and conservative, and design driving style recognition rules based on cluster centers. The process is as follows:
[0075] Step 5.1, initialize the SOM grid:
[0076] The SOM grid in this embodiment is a 1×3 self-organizing mapping (SOM) network. Each node j has a weight vector w j , with dimension k; initialize the weight vector w j ;
[0077] Step 5.2: Calculate the distance between the input sample and the weight vector:
[0078] For each input sample x i , calculate its Euclidean distance from the weight vector w j of each node j. The calculation formula for the Euclidean distance is d(x i , w j ) = ||x i - w j || 2 ;
[0079] Step 5.3: Find the best matching unit (BMU):
[0080] Find the node j i with the smallest distance from the input sample x * . The calculation formula is
[0081] Step 5.4: Update the weight vectors of BMUj * and its neighborhood nodes:
[0082] The formula is where t is the current iteration number, η(t) is the learning rate, which decays over time; is the neighborhood function, using a Gaussian function, and the formula is where and r j are the positions of the best matching unit (BMU) and node j in the grid respectively, and σ(t) is the neighborhood radius, which decays over time;
[0083] Step 5.5: Repeat steps 5.2 to 5.4 for 100 iterations until the maximum iteration number is reached or the weight vectors converge, train the SOM network, and obtain the clustering results (aggressive, stable, conservative).
[0084] Step 6: Analyze the SOM clustering method used in step 5 to improve its limitations; correct and optimize the SOM clustering method based on the Bayesian regularization method; design the improved driving style recognition rules; including the following steps:
[0085] Step 6.1: Optimize the SOM weight adjustment formula:
[0086] In the traditional SOM weight update formula, add a regularization penalty term, define the correction vector, and the new correction vector objective function is mk (t) = δE k (t)+∈E w (t), the modified weight adjustment formula, δ and ∈ are hyperparameters, and the hyperparameters δ and ∈ control the weights of the error term and the regularization term respectively; among them, the error term E k (t) is the difference between the input sample and the weight, E k (t) = x k -ω k (t); Regularization term E w (t) is the weighted mean square value, Used to penalize weight complexity, x k represents the kth eigenvalue of the input sample, ω k (t) represents the weight vector of neuron k at time step t, that is, the position of the neuron in the feature space; w ij (t) represents the connection weight from input node i to output node j at time step t, which is a specific element of the weight matrix in the SOM network;
[0087] Step 6.2, Bayesian reasoning dynamic optimization:
[0088] According to Bayesian reasoning, assuming that the hyperparameters δ and ∈ obey Gaussian distribution, the objective function m is solved by maximizing the posterior probability. k (t) The hyperparameters δ and ∈ at their minimum point, i.e. the optimal value of the hyperparameters Where, λ = n-2δ·Trace(T -1 ), T is the Hessian matrix of the modified objective function at the minimum point; the Bayesian method iteratively updates the hyperparameters to gradually make the weight distribution consistent with the probability distribution of the input data;
[0089] Step 6.3, update the SOM neuron position:
[0090] Taking the winning neuron as the benchmark, adjust the positions of other neurons during the iteration process: Among them, δ vk and d vk are the distances in high-dimensional and low-dimensional space respectively, p k (t) is the neuron position, α(t) is the learning rate, p v (t) is the position of the winning neuron.
[0091] Simulation experiment process example
[0092] 1. Data processing
[0093] The lane changing condition includes two parts of information, horizontal and vertical. Therefore, this embodiment selects 14 parameters, including the maximum absolute value of the steering wheel angle, lane changing time, lane changing initial time and front and rear vehicle distance, to describe the lane changing condition. The specific code of each parameter is shown in the table below.
[0094] Table 1 Statistical characteristic parameters
[0095]
[0096]
[0097] 2. Algorithm Implementation
[0098] The above clustering algorithm is implemented using Matlab, and the results of the two algorithms are compared, as shown in the following figure. Figure 2-4 As shown in the figure, the silhouette coefficient of the Bayesian improved SOM algorithm is larger and the clustering effect is better.
Claims
1. A driving style recognition method based on lane-changing driving data, characterized in that: The following steps are involved: Step 1: Build a driving data collection platform and construct a driver driving scenario; Design lane-changing scenarios under different working conditions; Step 2, collecting lane change data of drivers with different driving styles under different working conditions; The lane change data includes vehicle status data and vehicle operation data; Step 3: Collect driving data according to the start and end time of lane change; perform statistical processing on the raw data to obtain driver style characteristic parameters; Step 4: Use the principal component analysis method to reduce the dimension of the statistical features of each driving segment to obtain a new principal component after dimension reduction. The principal component is formed by a linear combination of various parameters in the index; then sort the principal components to obtain the principal component contribution rate and cumulative contribution rate, and extract the driving characteristics; Step 5: Cluster each driving feature using the SOM clustering method to form multiple driving style types, and design a driving style recognition rule based on the cluster center. The steps include: Step 5.1, initialize the SOM grid: The SOM grid is an m×m two-dimensional grid, and each node J has a weight vector w j , dimension is k; initialize the weight vector w j ; Step 5.2, calculate the distance between the input sample and the weight vector: For each input sample x i , calculate the weight vector w between it and each node j j The Euclidean distance is calculated as d(x i ,w j )=||x i -w j ||2; Step 5.3, find the best matching unit: Find the input sample x i The node j with the smallest distance * , the calculation formula is Step 5.
4. Update BMUj * and the weight vector of its neighboring nodes: The formula is Where t is the current iteration number, η(t) is the learning rate, which decays over time; is the neighborhood function, using the Gaussian function, the formula is in, and r j are the positions of the best matching unit and node j in the grid, respectively, σ(t) is the neighborhood radius, which decays with time; Step 5.5, iteration, repeat steps 5.2 to 5.4 until the maximum number of iterations is reached or the weight vector converges; Step 6: Analyze the SOM clustering method used in step 5 to improve its limitations; modify and optimize the SOM clustering method based on the Bayesian regularization method; design an improved driving style recognition rule; including the following steps: Step 6.1, optimize the SOM weight adjustment formula: In the traditional SOM weight update formula, a regularization penalty term is added and the new correction vector objective function is defined as m k (t) = δE k (t)+∈E w (t), the modified weight adjustment formula, δ and ∈ are hyperparameters, and the hyperparameters δ and ∈ control the weights of the error term and the regularization term respectively; among them, the error term E k (t) is the difference between the input sample and the weight, E k (t) = x k -ω k (t); Regularization term E w (t) is the weighted mean square value, x k represents the kth eigenvalue of the input sample, ω k (t) represents the weight vector of neuron k at time step t, that is, the position of the neuron in the feature space; w ij (t) represents the connection weight from input node i to output node j at time step t, which is a specific element of the weight matrix in the SOM network; Step 6.2, Bayesian reasoning dynamic optimization: According to Bayesian reasoning, the hyperparameters δ and ∈ obey Gaussian distribution, and the objective function m is solved by maximizing the posterior probability. k (t) Hyperparameters δ and ∈ at their minimum point, optimal value of hyperparameters Where, λ = n-2δ·Trace(T -1 ), T is the Hessian matrix of the modified objective function at the minimum point; the Bayesian method iteratively updates the hyperparameters to gradually make the weight distribution consistent with the probability distribution of the input data; Step 6.3, update the SOM neuron position: Taking the winning neuron as the benchmark, adjust the positions of other neurons during the iteration process: Among them, δ vk and d vk are the distances in high-dimensional and low-dimensional space respectively, p k (t) is the neuron position, α(t) is the learning rate, p v (t) is the position of the winning neuron.
2. The driving style recognition method based on lane-changing driving data according to claim 1, characterized in that: In the lane-changing scenario of step 1, the expected driving speed of the test vehicle is required to be within the range of 60km / h to 80km / h; there is a certain initial distance between the front vehicle and the test vehicle in the same lane, and the speed of the front vehicle is lower than the expected speed of the test vehicle; at the same time, there is a certain initial distance between the rear vehicle and the test vehicle in the other lane, and the speed of the rear vehicle is close to the expected speed of the test vehicle.
3. The driving style recognition method based on lane-changing driving data according to claim 1, characterized in that: The vehicle status data in step 2 includes one or more of the longitudinal displacement of the main vehicle, the lateral displacement of the main vehicle, the longitudinal speed of the main vehicle, the lateral speed of the main vehicle, the longitudinal acceleration of the main vehicle, and the lateral acceleration of the main vehicle; the vehicle control data includes one or two of the steering wheel angle of the main vehicle and the throttle opening of the main vehicle.
4. The driving style recognition method based on lane-changing driving data according to claim 1, characterized in that: Step 3 includes: Extract the lane-changing center time of the main vehicle; Extract the starting point time of the main vehicle changing lanes; Extract the lane-changing end point time of the main vehicle.
5. The driving style recognition method based on lane-changing driving data according to claim 1, characterized in that: The principal component analysis method described in step 4 comprises the following steps: Step 4.1: Data standardization: Suppose the original data set is X, where each row is a sample and each column is a feature; the standardized data is: Among them, μ is the mean vector of each feature, and σ is the standard deviation vector of each feature; Step 4.2, calculate the covariance matrix: The covariance matrix calculation formula is: Where n is the number of samples, ∑ is a d×d symmetric matrix, d is the number of features, X scaled To standardize the data set; Step 4.3, select the main components: Select the eigenvectors corresponding to the k largest eigenvalues to form the projection matrix W; W = Q :k ; Among them, Q :k is the first k columns of Q; Step 4.4, data projection: Project the standardized data into the principal component space, the formula is X pca =X scaled W, where X pca is the data after dimensionality reduction, with a dimension of n×k.