A driving lane-changing style recognition method based on improved particle swarm optimization SVM

Abstract

The application discloses a driving lane-changing style recognition method based on an improved particle swarm optimization SVM, which comprises the following steps: S1, extracting data as trajectory data in the vehicle driving process based on the NGSIM data set; S2, performing preliminary screening on the extracted driver lane-changing data, and extracting lane-changing characteristic parameters of the driver lane-changing style; S3, using a Kmeans clustering algorithm to cluster the extracted lane-changing data into three categories (cautious type, general type and aggressive type); S4, performing statistical analysis and time-frequency analysis on the clustered parameters to verify the difference of the classified drivers, and extracting effective parameters as inputs of a recognition model; and S5, improving the particle swarm optimization algorithm based on the strategy of nonlinear change of inertia weight and nonlinear decrease of maximum particle speed, updating two key parameters C and g of the SVM model, and establishing a driver style recognition model based on the improved particle swarm optimization SVM, so as to recognize the driver lane-changing style.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent driving or driving style recognition, and particularly relates to a driving lane-changing style recognition method based on an improved particle swarm optimization support vector machine. BACKGROUND

[0002] Driving is a complex activity, and its safety is affected by different factors such as driver behavior, environmental conditions, and the vehicle itself. The development of advanced driving assistance systems (ADAS) and other active safety technologies is particularly rapid, and the next frontier of lane-changing collision prevention will be the design of scene-based technology. Due to the differences in the psychological state, personality, and physical function of each driver, the driving styles of different drivers will also show significant differences. However, the existing ADAS systems rarely have individual control strategies for different drivers. Therefore, in order to meet the different needs of ADAS users, it is necessary to provide adaptive regulation of different driver behavior characteristics to meet different requirements.

[0003] Particle swarm optimization (PSO) is a kind of random optimization technology based on grouping, and its basic concept comes from the study of bird foraging behavior. Through the particle swarm optimization, the two key parameters C and g of the support vector machine (SVM) can be updated and optimized to provide recognition rate, but the standard PSO algorithm is prone to local optimization and has slow convergence speed. Based on the strategy of nonlinear change of inertia weight and nonlinear decrease of maximum particle speed, the particle swarm is improved, and the support vector machine optimized by the improved particle swarm algorithm is used to recognize the lane-changing parameters of the driver, thereby improving the recognition accuracy. SUMMARY

[0004] The purpose of the present application is to provide a driving lane-changing style recognition method based on an improved particle swarm optimization support vector machine. Based on the NGSIM data set, effective natural driving lane-changing trajectory data is extracted, the drivers are divided into three categories, and the differences between the three categories of drivers are analyzed through statistical analysis and time-frequency analysis. The lane-changing parameter indicators are extracted, and the support vector machine optimized by the improved particle swarm algorithm is used to recognize the lane-changing parameters of the driver.

[0005] The purpose of the present application is achieved by the following technical solutions:

[0006] A driving lane-changing style recognition method based on an improved particle swarm optimization support vector machine, comprising the following steps:

[0007] S1, the experiment is based on the data extracted from the NGSIM data set, which includes trajectory data and surrounding environment data during vehicle driving, specifically including: lane-changing time, speed, acceleration, following distance, relative distance to the front vehicle, etc.

[0008] S2, the extracted driver lane changing data is preliminarily screened, lane changing data with small influence degree is eliminated, and lane changing characteristic parameters reflecting the lane changing style of the driver are reserved.

[0009] S3, the lane changing data extracted is clustered into three categories (cautious type, general type, and aggressive type) by using Kmeans clustering algorithm.

[0010] S4, and the statistical analysis and time-frequency analysis of the clustered parameters are used to verify the difference of the classified drivers, and the effective parameters are extracted as the input of the recognition model.

[0011] S5, based on the strategy of nonlinear change of inertia weight and nonlinear decrease of maximum particle speed, the particle swarm optimization algorithm is improved, the two key parameters C and g of the SVM model are optimized and updated, and the driver style recognition model based on improved particle swarm optimization SVM is established, and the lane changing style of the driver is recognized.

[0012] Further, the lane changing data extracted is clustered by using Kmeans clustering algorithm:

[0013] S3.1, the lane changing characteristic parameters for clustering this time are lane changing time, speed, acceleration, following time distance, relative distance to the front vehicle, and lateral displacement difference, and three categories are set, which are cautious type, general type, and aggressive type.

[0014] S3.2, the initial clustering center is randomly selected, and the original variables are clustered into 3 categories according to the principle of minimum square of Euclidean distance.

[0015] S3.3, the average value of each category is continuously calculated to obtain the clustering calculation until the termination iteration condition is reached to determine the new clustering center.

[0016] Further, the statistical analysis and time-frequency analysis of the clustered parameters are used to verify the difference of the classified drivers

[0017] S4.1, statistical analysis is performed on the lane changing characteristic parameters

[0018] The main processing is the mean value of lane changing time, speed, acceleration, following time distance, and relative distance to the front vehicle, which is represented as:

[0019] Standard deviation, represented as: Maximum value: X max = max(x i ), minimum value: X min = min(x i ), wherein x i is the selected sample. Among them, the variance test is performed on the obvious difference, and the K-W test is used for the non-significant difference:

[0020] Where: N is the sample size; R j It is the rank sum; n j J is its measured value; J is the number of sample groups.

[0021] S4.2. Perform time-frequency analysis on lane-changing characteristic parameters.

[0022] By using wavelet transform to perform localized analysis of the time (space) frequency of a signal (function), the characteristic details of different frequencies are highlighted, thereby achieving the division of high and low frequencies and completing time-frequency analysis to analyze the differences in drivers' lane-changing styles.

[0023] If the signal is f(t) and r0 is the starting point of the wavelet being measured, then its series expansion is as follows:

[0024]

[0025] Where: r is the scale; t is time; s is displacement. It is an approximation coefficient, d r (s) is the detail coefficient. It is a scaling function, ψ r,s (t) is a wavelet function. When these two functions are orthogonal, then and d r It can be represented as:

[0026]

[0027] Where <.> represents the inner product calculation method.

[0028] Furthermore, the process of optimizing and updating the two key parameters C and g of the support vector machine using the improved particle swarm optimization algorithm is as follows:

[0029] S5.1 When using the SVM method to find the optimal classification function on the training set, the optimal function is transformed into the problem of finding the maximum classification interval between hyperplanes. Given a sample training set D = {(x...} i ,y i ),i=1,2,…,l}∈(R n ×R), to establish the relevant linear functions in the high-dimensional space;

[0030] f(x) = ωΦ(x) + b

[0031] Furthermore, the classification problem can be transformed into the following equation for solution:

[0032]

[0033] Where ω is the weight vector and b is the offset vector, and the training sample is (xi, yi), ξ i ,ξ * iLet represent the slack variable, and n represent the number of training samples. Parameter C is an adjustable parameter, also known as the penalty coefficient.

[0034] To solve the above equation, the problem can be transformed using the Lagrange multiplier method:

[0035]

[0036] Where α is a Lagrange multiplier and K(xi,yi) is the kernel function. Because radial basis functions have a wide region of convergence, they can be chosen as the kernel function, expressed as:

[0037]

[0038] parameters Substituting into the above formula, we get:

[0039]

[0040] In the process of building a Support Vector Machine (SVM) model, the selection of parameters C and kernel parameters g is crucial to the algorithm's performance. Currently, there is no explicit definition for the selection of parameters C and g. Therefore, the Particle Swarm Optimization (PSO) algorithm is used to find the optimal parameters to obtain the optimal SVM model.

[0041] S5.2, Particle Swarm Optimization Algorithm

[0042] The PSO algorithm, a group-based stochastic optimization technique, is based on the study of bird foraging behavior. The i-th particle is denoted as X. i =(x i1 ,x i2 , ..., x iD Its optimal position is P. i =(p i1 ,p i2 , ..., p iD ), that is, P Best The subscript of the optimal position experienced by each particle in the population is denoted by the symbol g, where g is p. g The velocity of particle i is expressed as V. i =(v i1 ,v i2 , ..., v iD For each generation, its d-th dimension (1≤d≤D) updates its velocity and new position according to the following formula:

[0043] v id =ωv id +c1rand()(p id -x id )+c2Rand()(p gd -x id )

[0044] x id =x id +v id , where ω is the inertia weight, c1 and c2 are learning factors, and rand() and Rand() are two random functions that vary between (0,1).

[0045] S5.3. Inertia weight has a significant impact on the optimization ability of particles. Therefore, based on the nonlinear variation strategy of inertia weight, the size of the weight factor is controlled to control the proportion of the maximum weight, so that the particle swarm can perform more accurate initial global search and final local search. This can be expressed as:

[0046] Where t is the iteration number, T is the total number of iterations, and K is the weight factor.

[0047] Excessive particle velocity and insufficient velocity can lead to failure to find the global optimum and local optimum, respectively. The method based on the nonlinear decrease of the particle's maximum velocity helps reduce the loss of particles in the search region due to excessive velocity, thus reducing invalid searches. This can be expressed as:

[0048] Where V max With V min It represents the maximum and minimum values ​​of the maximum speed.

[0049] S5.4 The concept of the improved PSO-SVM recognition model is to optimize two key parameters using an improved particle swarm algorithm. Based on the data sample set, we randomly select about 70% of the data as the training set and 30% of the data as the test set, and then train the model.

[0050] S5.4.1 Initialize the velocity and position of each particle in the search space, calculate the fitness function value, and obtain the historical best position of the particles and the global best position of the swarm.

[0051] S5.4.2 Update of Particle Velocity and Position: Each particle's velocity and position are updated based on its historical best position and the global best position. Here, a non-linear inertia weight and a non-linear decreasing strategy for maximum particle velocity are added to the particle swarm algorithm. This allows the particle swarm to perform a more accurate initial global search and a final local search, while reducing the loss of particles in the search area due to excessive velocity and minimizing invalid searches.

[0052] S5.4.3. Evaluate the fitness value of each particle, update the historical best position and the global best position of the particles. Check if the iteration meets the termination condition.

[0053] S5.4.4: Set the maximum number of iterations for the termination condition. If the condition is met, output the optimal parameters C and g, input training samples and prediction samples, normalize the data, and establish an improved SVM model for particle swarm optimization. Otherwise, repeat S5.4.2.

[0054] Furthermore, after establishing the model, the three clustered driver types are used as the output of the model, and the analyzed lane-changing feature parameters are extracted as input to verify the recognition effect of the model.

[0055] The beneficial effects of this invention are as follows:

[0056] 1. The purpose of this invention is to propose a driving lane-changing style identification method based on an improved particle swarm optimization algorithm to optimize the support vector machine. Its beneficial effect is that the improved particle swarm optimization algorithm is used to optimize two key parameters of the support vector machine algorithm. The optimized SVM model identifies the driver's lane-changing style. Compared with the traditional particle swarm optimization algorithm, the improved particle swarm optimization algorithm has better convergence accuracy and convergence speed, which is beneficial for local and global search.

[0057] 2. This invention proposes to use statistical analysis and time-frequency analysis to analyze the differences in driver lane-changing parameters, which is helpful to better describe the differences in drivers during the lane-changing process.

[0058] 3. This invention provides a new method in the field of driver identification, which has good application prospects in the future fields of intelligent driving and driver assistance systems. Attached Figure Description

[0059] Figure 1 This is a flowchart of the driver lane-changing style identification method of the present invention;

[0060] Figure 2 This is a flowchart of the driver clustering method of the present invention;

[0061] Figure 3 This is a flowchart of the wavelet analysis method of the present invention;

[0062] Figure 4 This is a schematic diagram of the improved particle swarm iteration curve of the present invention;

[0063] Figure 5 This is a flowchart of the proposed algorithm for optimizing the support vector machine based on the improved particle swarm optimization algorithm. Detailed Implementation

[0064] To better illustrate the objectives and technical solutions of this invention, the 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 merely illustrative of the invention and not intended to limit it. Furthermore, for ease of description, the accompanying drawings show only the parts relevant to the invention and not the entire structure.

[0065] As attached Figure 1 The diagram illustrates the operational flow of the proposed driver lane-changing style identification method based on an improved particle swarm optimization algorithm and optimized support vector machine. Valid lane-changing data is extracted from the NGSIM dataset, with unreasonable data being filtered out. The extracted valid data is processed and analyzed to obtain feature parameters. The K-means algorithm is used to cluster drivers, classifying them into three categories: cautious, general, and aggressive. Statistical and time-frequency analyses are performed on the clustering parameters to verify the differences in driver classification, and valid parameters are extracted as input to the identification model. Then, a strategy based on nonlinear changes in inertial weights and nonlinear decreases in maximum particle velocity is used to improve the particle swarm optimization. Finally, a driver style identification model based on the improved particle swarm optimization and optimized SVM is established. The two key parameters C and g of the SVM model are optimized and updated, achieving a driver identification accuracy of 96.25%.

[0066] The specific implementation process of this invention is as follows:

[0067] S1. Based on the NGSIM dataset, a total of 532 sets of valid data were obtained using the PYTHON data processing library. The lane-changing feature indicators extracted in the experiment are the trajectory data and surrounding environment data during vehicle driving, specifically including: lane-changing time, speed, acceleration, following distance, and relative distance to the vehicle in front.

[0068]

[0069] S2. Filter and extract vehicle data of type 2, namely small cars, that have lane-changing behavior, and remove unreasonable driving data that exceeds the prescribed road or speed.

[0070] S3. Based on the feature parameters in S1, form a sample set A = {x1, x2, ..., x...} 532 There are a total of 532 lane-change data sets, with each driver having a sample size of x. i =(x i1, x i2 ,…,x i6 ) TAs a six-dimensional feature vector including lane-change time, speed, acceleration, following distance, and relative distance to the vehicle in front, the extracted lane-change data sample set A is clustered into three categories (cautious, general, and aggressive) using the K-means clustering algorithm. The K-means clustering algorithm is written in MATLAB, and its process is as follows: Figure 2 As shown.

[0071] S3.1. The 532 sets of driver data collected after processing are used as a sample set A = {x1, x2, ..., x...} 532}, each driver sample x i =(x i1, x i2 ,…,x i6 ) T As a six-dimensional feature vector, it includes lane change time, speed, acceleration, following distance, and relative distance to the vehicle in front, and the cluster is set to k=3;

[0072] S3.2. The 532 sets of samples are assigned to the clustering algorithm. According to the algorithm principle, the original total number of variables n is clustered into k classes based on the principle of minimizing the square of the Euclidean distance.

[0073] S3.3. New cluster centers can be determined by continuously calculating the average value of each class and then performing cluster calculations until the termination iteration condition is met.

[0074] S4. Statistical and time-frequency analyses are performed on the clustering parameters to verify the differences in driver classification, and effective parameters are extracted as input to the identification model.

[0075] S4.1 Statistical analysis of lane-changing characteristic parameters

[0076] The main parameters processed are the average values ​​of lane-changing time, speed, acceleration, following distance, and relative distance to the vehicle in front, expressed as:

[0077] Standard deviation, expressed as: Maximum value: X max =max(x i Minimum value: X min =min(x i ), where x i For the selected sample, a variance test was performed on those with significant differences, and a Karl von W. test was used on those without significant differences.

[0078] Where: N is the sample size; R j It is the rank sum; n j J is its measured value; J is the number of sample groups.

[0079] The K-means algorithm is used to cluster the original total number of variables n into k classes based on the principle of minimizing the square of the Euclidean distance. The new cluster centers can be determined by continuously calculating the average value of each class and then performing clustering calculations to set the clusters into three classes, and then determining the cluster centers based on the characteristics of the clustering results.

[0080] The mean values ​​(within the 95% confidence interval) for lane change time, speed, acceleration, following distance, and relative distance to the vehicle in front are shown in the table below:

[0081]

[0082] The KW test results showed that the median Sig values ​​for v and x0 were 0.023 and 0.532, respectively, both greater than 0.01, indicating no significant difference. All other indicators were less than 0.01, suggesting a difference.

[0083] S4.2. Perform time-frequency analysis on lane-changing characteristic parameters.

[0084] By utilizing wavelet transform for time (space) frequency localization analysis of a signal (function), the characteristic details of different frequencies are highlighted, thereby achieving the distinction between high and low frequencies and completing time-frequency analysis. This allows for the analysis of differences in driver lane-changing styles. The calculation process is as follows: Figure 3 As shown.

[0085] S5. Based on the strategy of nonlinear change of inertia weight and nonlinear decrease of particle maximum velocity, the particle swarm algorithm is improved. The two key parameters C and g of the SVM model are optimized and updated. A driver style recognition model based on the improved particle swarm optimization SVM is established to identify the driver's lane-changing style.

[0086] S5.1 First, establish the SVM model, which is a linear function in the high-dimensional space;

[0087] f(x) = ωΦ(x) + b

[0088] Furthermore, the classification problem can be transformed into the following equation for solution:

[0089]

[0090] Where ω is the weight vector and b is the offset vector, and the training sample is (xi, yi), ξ i ,ξ * i Let represent the slack variable, and n represent the number of training samples. Parameter C is an adjustable parameter, also known as the penalty coefficient.

[0091] To solve the above equation, the problem can be transformed using the Lagrange multiplier method:

[0092]

[0093] Where α is a Lagrange multiplier and K(xi,yi) is the kernel function. Because radial basis functions have a wide region of convergence, they can be chosen as the kernel function, expressed as:

[0094]

[0095] parameters Substituting into the above formula, we get:

[0096]

[0097] Furthermore, the two parameters are optimized using an improved particle swarm optimization algorithm.

[0098] S5.2, Particle Swarm Optimization Algorithm

[0099] The PSO algorithm, a group-based stochastic optimization technique, is based on the study of bird foraging behavior. The i-th particle is denoted as X. i =(x i1 ,x i2 , ..., x iD Its optimal position is P. i =(p i1 ,p i2 , ..., p iD ), that is, P Best The subscript of the optimal position experienced by each particle in the population is denoted by the symbol g, where g is p. g The velocity of particle i is expressed as V. i =(v i1 ,v i2 , ..., v iD For each generation, its d-th dimension (1≤d≤D) updates its velocity and new position according to the following formula:

[0100] v id =ωv id +c1rand()(p id -x id )+c2Rand()(p gd -x id )

[0101] x id =x id +v id , where ω is the inertia weight, c1 and c2 are learning factors, and rand() and Rand() are two random functions that vary between (0,1).

[0102] S5.3, such as Figure 4 The following two optimal parameters can be obtained using an improved particle swarm optimization algorithm:

[0103] S5.3.1. Employing a nonlinear variation strategy for inertial weights, the proportion of the maximum weight is controlled by adjusting the magnitude of the weighting factor. This can be expressed as:

[0104] Where t is the iteration number, T is the total number of iterations, and K is the weight factor.

[0105] S5.3.1, The nonlinear decreasing strategy based on the maximum velocity of particles helps to reduce the loss of particles in the search area due to excessive velocity and reduces invalid searches, which can be expressed as:

[0106] Where V max With V min It represents the maximum and minimum values ​​of the maximum speed.

[0107] S5.2, such as Figure 5 As shown, the process of improving the particle swarm optimization algorithm to optimize the SVM driving style recognition model is as follows: First, different types of drivers and their corresponding feature parameters are used as the output and input of the model, respectively. The data processing is saved as a MATLAB .mat numerical matrix file. A support vector machine model is established, and the support vector machine model is optimized using the improved particle swarm optimization algorithm. The two optimal parameters are output. 70% of the data is used as the training set and 30% of the data is used as the test set. Then, the model is trained to recognize lane-changing styles.

Claims

1. A driving lane-changing style recognition method based on an improved particle swarm optimization algorithm-optimized support vector machine, characterized in that, Includes the following steps: S1. The data extracted from the NGSIM dataset in the experiment are trajectory data and surrounding environment data during vehicle driving, specifically including: lane change time, speed, acceleration, following distance, and relative distance to the vehicle in front; S2. Perform preliminary screening on the extracted driver lane change data, remove lane change data with little impact on lane changes, and retain lane change characteristic parameters that can reflect the driver's lane change style. S3. Use the K-means clustering algorithm to cluster the extracted lane-changing data into three categories: cautious, general, and aggressive. S4. Statistical and time-frequency analysis of the clustering parameters are performed to verify the differences in the classification of drivers, and effective parameters are extracted as input to the identification model. S5. Based on the strategy of nonlinear change of inertia weight and nonlinear decrease of particle maximum velocity, the particle swarm algorithm is improved. The two key parameters C and g of the SVM model are optimized and updated. A driver style recognition model based on the improved particle swarm optimization SVM is established to identify the driver's lane changing style. The specific process of step S4 is as follows: S4.1 Statistical analysis of lane-changing characteristic parameters The data is processed into the mean of lane-changing time, speed, acceleration, following distance, and relative distance to the vehicle in front, expressed as the sample mean: Where: i represents a single sample; x i For each sample parameter of the lane change, n is the total number of samples; the standard deviation is expressed as: Maximum value: X max =max(x i Minimum value: X min =min(x i ), where x i For the selected samples, a variance test was performed on those with significant differences, and a KW test was used on those without significant differences. Where: N is the sample size; R j It is the rank sum; n j It is its measured value; J is the number of sample groups; S4.

2. Perform time-frequency analysis on lane-changing characteristic parameters. By using wavelet transform to perform localized analysis of the time (space) frequency of a signal (function), the characteristic details of different frequencies are highlighted to achieve the division of high and low frequencies, thereby completing time-frequency analysis and analyzing the differences in drivers' lane-changing styles. If the signal is f(t) and r0 is the starting point of the wavelet being measured, then its series expansion is as follows: Where: r is the scale; t is time; s is displacement. It is an approximation coefficient, d r (s) is the detail coefficient. It is a scaling function, ψ r，s (t) is a wavelet function. When these two functions are orthogonal, then and d r It can be represented as: Where <·> represents the inner product calculation method.

2. The driving lane-changing style recognition method based on an improved particle swarm optimization algorithm for support vector machines according to claim 1, characterized in that, The specific process of step S3 is as follows: S3.1 The lane-changing feature parameters used for clustering are lane-changing time, speed, acceleration, following distance, relative distance to the vehicle in front, and lateral displacement difference. The K-means algorithm is used to cluster the original total number of variables n into k classes based on the principle of minimizing the square of the Euclidean distance. The new cluster centers can be determined by continuously calculating the average value of each class and then performing clustering calculations to set three classes. Based on the characteristics of the clustering results, the three types are determined to be cautious, general, and aggressive. S3.2 Randomly select initial cluster centers and cluster the original variables into 3 classes based on the principle of minimizing the square of the Euclidean distance; S3.3 Continuously calculate the average value of each class to obtain the clustering calculation until the termination iteration condition is reached to determine the new cluster center.

3. The driving lane-changing style recognition method based on an improved particle swarm optimization algorithm-optimized support vector machine as described in claim 1, characterized in that, In step S5, the process of optimizing and updating the two key parameters C and g of the support vector machine using the improved particle swarm optimization algorithm is as follows: S5.1 When using the SVM method to find the optimal classification function on the training set, the optimal function is transformed into the problem of finding the maximum classification interval between hyperplanes. Given a sample training set D = {(x i ,y i ),i=1,2,…,l}∈(R n ×R), to establish the relevant linear functions in the high-dimensional space; f(x) = ωΦ(x) + b Furthermore, the classification problem can be transformed into the following equation for solution: Where y i Let ω be the ordinate of the corresponding training sample, b be the weight vector, and b be the offset vector. The training sample is (x... i ,y i ), ξ i ,ξ * i Let ε represent the slack variable, ε be the threshold, n be the number of training samples, and C be an adjustable parameter, also known as the penalty coefficient. To solve the above equation, the problem can be transformed using the Lagrange multiplier method: Where α is a Lagrange multiplier and K(xi,yi) is the kernel function. Since the radial basis function has a wide region of convergence, it can be chosen as the kernel function, expressed as: parameters Substituting into the above formula, we get: In the process of building a support vector machine model, the selection of parameters C and kernel parameters g is crucial to the performance of the algorithm. Currently, there is no clear definition for the selection of parameters C and kernel parameters g. Therefore, the particle swarm optimization algorithm is used to find the optimal parameters and obtain the optimal SVM model. S5.2, Particle Swarm Optimization Algorithm The PSO algorithm is a group-based stochastic optimization technique, where the i-th particle is represented as X. i =(x i1 ,x i2 , ..., x iD Its optimal position is P. i =(p i1 ,p i2 , ..., p iD ), that is, P Best The subscript of the optimal position experienced by each particle in the population is denoted by the symbol g, where g is p g The velocity of particle i is expressed as V i =(v i1 ,v i2 , ..., v iD For each generation, its d-th dimension (1≤d≤D) updates its velocity and new position according to the following formula: v id =ωv id +c1rand()(p id -x id )+c2Rand()(p gd -x id ) x id =x id +v id , where ω is the inertia weight, c1 and c2 are learning factors, and rand() and Rand() are two random functions that vary between (0,1); S5.

3. Inertia weight has a significant impact on the optimization ability of particles. Therefore, based on the nonlinear variation strategy of inertia weight, the size of the weight factor is controlled to control the proportion of the maximum weight, so that the particle swarm can perform more accurate initial global search and final local search. This can be expressed as: Where t is the iteration number, T is the total number of iterations, and K is the weight factor. Excessive particle velocity and insufficient velocity can lead to failure to find the global optimum and local optimum, respectively. The method based on the nonlinear decrease of the maximum particle velocity helps reduce the loss of particles in the search region due to excessive velocity, thus reducing ineffective searches. The maximum particle velocity changing over time can be expressed as: Where t is the iteration number, T is the total number of iterations, and v max With v min These are the maximum and minimum values ​​of the maximum speed; S5.4 The concept of improving the PSO-SVM recognition model is to optimize two key parameters using an improved particle swarm algorithm. Based on the data sample set, we randomly select about 70% of the data as the training set and 30% of the data as the test set, and then train the model. S5.4.1 Initialize the velocity and position of each particle in the search space, calculate the fitness function value, and obtain the historical best position of the particles and the global best position of the swarm. S5.4.2 Update of particle velocity and position: Based on its own historical best position and global best position, the velocity and position of each particle are updated. Here, inertial weight nonlinearity and particle maximum velocity nonlinearity decreasing strategy are added to the particle swarm algorithm, so that the particle swarm can perform more accurate initial global search and final local search, and reduce the loss of particles in the search area due to excessive velocity, and reduce invalid search. S5.4.

3. Evaluate the fitness value of each particle, update the historical best position and the global best position of the particles, and check whether the iteration meets the termination condition. S5.4.4 Set the maximum number of iterations for the termination condition. If the condition is met, output the optimal parameters C and g, input the training samples and prediction samples, normalize the data, and establish an improved particle swarm optimization SVM model. Otherwise, repeat S5.4.

2.

4. The driving lane-changing style recognition method based on an improved particle swarm optimization algorithm for support vector machines according to claim 1, characterized in that, It also includes establishing a model, using the three clustered driver types as the model's output, and extracting and analyzing lane-changing feature parameters as input to verify the model's recognition performance.

Images