Improved acc control method based on preceding vehicle driving style recognition

Abstract

The application relates to an improved ACC control method based on front vehicle driving style recognition, which is characterized by the following specific implementation steps: 1, selecting a vehicle trajectory data set as training data, first, pre-processing the data, 2, adopting FMC fuzzy C means clustering to classify the processed sample data, 3, establishing an evaluation index of the clustering result, so as to determine the dimension of the optimal feature vector and the number of clustering clusters: 4: establishing a follow-up model of the adaptive cruise control, the follow-up model is characterized by correcting the follow-up distance based on the driving style of the front vehicle, the follow-up strategy of the adaptive cruise system is adjusted through the recognition of the driving style of the front vehicle, the driving efficiency and the utilization rate of the road can be improved under the premise of ensuring safety, and the application has a positive effect on improving the traffic condition and saving energy.

Description

Technical Field

[0001] This invention relates to an improved ACC control method based on the recognition of the driving style of the vehicle in front, belonging to the field of cruise control following models. Background Technology

[0002] Most existing studies on driver style classification and identification focus on drivers of their own vehicles. Therefore, there are many types of data used to determine driver style. Commonly used data includes the driver's pressure on the brake and drive pedals. Some studies also use sensors such as in-vehicle cameras or eye trackers to collect physiological signals such as the driver's emotions, attention, or heart rate to analyze the driver's style. However, in studies on the classification of driving styles of vehicles in front, the above studies are difficult to apply due to the lack of available data types. Summary of the Invention

[0003] The purpose of this invention is to provide an improved ACC control method based on the recognition of the driving style of the preceding vehicle. The ACC adaptive cruise control strategy of the vehicle is adjusted according to the different driving styles of the preceding vehicle. Under certain circumstances, the control strategy is adjusted to be more aggressive while ensuring safety, which can improve the driving efficiency of the ACC system. Because it requires fewer sensor signals, it can be applied to the classification of the driving style of the preceding vehicle.

[0004] The technical solution of this invention is implemented as follows: An improved ACC control method based on the driving style recognition of the preceding vehicle is a control method that uses FMC (fuzzy C-means clustering) to classify the driving style of longitudinally traveling vehicles on a highway according to their motion characteristics, thereby adjusting the adaptive cruise following strategy. The method is characterized by the following specific steps:

[0005] Step one: Select the vehicle trajectory dataset as training data and preprocess the data. This step mainly involves the selection of feature vectors, which is based on the adaptive cruise control's driving style.

[0006] 1) Slice the data and extract the segments of each vehicle's data that do not involve lane changing or turning. Perform isochronous processing on the data. Select the first 100 frames of motion for each data set and extract the mean of the velocity, the mean of the acceleration, and the variance as feature quantities.

[0007] 2) To avoid the influence of different evaluation metrics having different units, the data is normalized. First, a bar chart is plotted to show the distribution of the extracted data. A large amount of data is concentrated in a small subset of the entire value range, and a few outliers affect the maximum and minimum values ​​of the samples. Therefore, Z-Score standardization is used to preprocess the data to avoid the influence of outliers and to normalize the data to remove the influence of units. The Z-Score formula is as follows:

[0008] z=(x-μ) / σ (1)

[0009] Where x is the original score, μ is the mean score of the overall sample space, and σ is the standard deviation of the overall sample space.

[0010] Step two involves classifying the processed sample data. In this process, the FMC soft clustering method is used for style classification. Compared to the hard clustering methods, primarily K-means clustering, which are mainly used in existing research, the advantage of soft clustering is that the membership degree of each cluster is not simply either "belongs" or "does not belong," but can simultaneously belong to multiple clusters. The membership degree of a sample point to each cluster can be represented by the membership degree, which ranges from 0 to 1 to describe the degree of membership. This fuzzy clustering is more suitable for applications involving driving style classification. Since the transition from conservative to aggressive driving style does not have a clear dividing line and is a gradual transition, FMC fuzzy C-means clustering is used to classify the processed sample data, as detailed below:

[0011] First, an m-dimensional evaluation index is selected, and feature vectors X1, X2, ..., X3 are obtained from n sets of preprocessed data. n The sample dataset is then D = {X1, X2, ..., X...} n}, where sample X i ={x1,x2,...,x m Let} be an m-dimensional vector; given a predetermined number of categories c, then C1, C2, ..., C... C Given c clusters obtained by clustering dataset D, u i (i = 1, 2, ..., c) are the cluster centers of the i-th cluster, and a single sample X in D is given. k For the i-th cluster C i The membership degree is α ik (i=1,2,...,c;k=1,2,...,n),α ik The membership matrix U that constitutes c*n must satisfy the following conditions:

[0012] α ik ∈[0,1](1≤i≤c,1≤k≤n) (2)

[0013]

[0014] The following weighted distance is used to reflect the metric sample X. k and the i-th cluster C i The correlation between them is expressed by the following formula:

[0015]

[0016] Where u ij For the i-th cluster C i The j-th component of the cluster center, where m is the number of components of the feature vector.

[0017] Therefore, the sum of the distances of all samples to all clusters is:

[0018]

[0019] To adjust the influence of membership degree on distance calculation, the above formula is changed to a weighted sum of distances:

[0020]

[0021] In the formula, p > 1 is a fuzzy parameter, which determines the fuzziness of the clustering, that is, the degree to which a sample point can belong to multiple clusters. The optimal solution is when the weighted distance reaches its minimum value, and Equation (2) is used as a constraint condition for the membership degree, transforming the problem into a constrained minimum value problem. Therefore, the Lagrange multiplier method is used to solve the above problem, and the Lagrange function is:

[0022]

[0023] Where λ1, λ2, ..., λ k It is a Lagrange operator, x kj It is the j-th variable of the k-th sample, u ij It is the j-th variable of the i-th cluster center.

[0024] Let F(α) ik ,λ1,λ2,...,λ k For each α ik , λ k Taking the derivative, we have:

[0025]

[0026]

[0027] From equation (9), we can obtain:

[0028]

[0029] Substitute it And by organizing the information, we can finally arrive at the following solution:

[0030]

[0031] To find the optimal cluster centers, the objective function F(α) is used. ik ,λ1,λ2,...,λk ) is regarded as the cluster center u ij The function, i.e., F(u) ij By taking its derivative, the optimal cluster center coordinates u of each cluster can be calculated. i as follows:

[0032]

[0033] The formula for calculating the optimal cluster centers is as follows:

[0034]

[0035] Therefore, by using an iterative method to cluster the data and obtain the optimal clustering result, we can obtain the optimal membership matrix U and the optimal cluster centers u. j The specific steps are as follows:

[0036] 1) Determine the number of clusters, determine the initial parameter p, specify the allowable relative error, i.e., U(t)-U(t-1)<ζ, or specify the maximum number of iterations T;

[0037] 2) Randomly select the initial cluster centers u (0) Calculate the initial membership degree α using the above formula. (0) ;

[0038] 3) Calculate u again (1) ,α (1) The iteration stops when the relative error is less than the allowable error ζ or the number of iterations reaches the specified number.

[0039] Step 3: Establish evaluation metrics for clustering results to determine the optimal dimension of the feature vector and the number of clusters.

[0040] Because it is impossible to obtain an absolutely accurate classification of driving styles, traditional external evaluation metrics cannot be used for evaluation.

[0041] Using improved external metrics, three-dimensional eigenvectors (mean velocity, mean absolute acceleration, and standard deviation of acceleration) and two-dimensional eigenvectors (mean absolute acceleration and standard deviation of acceleration) were selected for binary and ternary clustering, respectively. The clustering effects of a total of four clustering methods were compared.

[0042] Using the headway as the criterion for category labeling, two evaluation methods are employed:

[0043] 1) Evaluation was performed using the improved adjusted Rand coefficient, with the following specific steps:

[0044] First, calculate the Rand coefficient RI:

[0045]

[0046] Here, TP represents the number of data points that belong to the same category in both external and internal metrics; TN represents the number of data points that do not belong to the same category in either external or internal metrics; FP represents the number of data points that do not belong to the same category in external metrics but belong to the same category in internal metrics; and FN represents the number of data points that belong to the same category in external metrics but do not belong to the same category in internal metrics. The four types of data point pairs in the denominator encompass all possible outcomes for any two randomly selected data points in the entire sample, thus simplifying the formula to:

[0047]

[0048] Clearly, TP and TN are correct decisions, while FP and FN are incorrect decisions. The RI value ranges from [0,1]. A drawback of RI is that it cannot make the value close to 0 in the case of completely random clustering. Therefore, the adjusted Rand coefficient (ARI) is used. The formula for calculating ARI from RI is as follows:

[0049]

[0050] 2) Visualize and compare using the headway distribution map method. The specific steps are as follows:

[0051] Since the average vehicle headway distance of vehicles in each driving style obtained by clustering is distributed in various intervals, the proportion of different average vehicle headway distances in each style under each clustering method is extracted and a bar chart is drawn, which is the vehicle headway distance distribution map of each style. Then, several sets of vehicle headway distance distribution maps of the same style under different clustering conditions are drawn into one map for comparison.

[0052] For the conservative style clustering, the average headway is used as the criterion. There should be more vehicles distributed in the high headway portion. Therefore, the more accurate headway distribution of the conservative clustering should be mainly distributed in the part with higher headway.

[0053] For the more aggressive type, the more accurate aggressive clustering method should mainly focus on the part with lower headway distance.

[0054] The height of the vehicle's front end is determined by the 2-second limit mentioned above, which was obtained by consulting regulations.

[0055] We selected a three-dimensional, three-class classification clustering method.

[0056] Step four: Establish a following model for adaptive cruise control. This following model is characterized by adjusting the following distance based on the driving style of the vehicle in front, as detailed below:

[0057] 1) First, establish a variable-distance following model:

[0058] d e =τ×v+d0 (17)

[0059] In the formula, d e For the desired following distance, τ is a coefficient related to the headway, which can be expressed as τ(t), v is the vehicle speed, and d0 is the minimum following distance. The safe following distance is calculated using the variable headway method, that is, τ is changed according to different driving conditions to adjust the following distance. The adjustment of τ is based on the driving style classification mentioned above and must meet the following constraints:

[0060] d f -d r +D≥d e (18)

[0061] Where d f d is the distance the car in front travels during the process of both cars reaching the same speed. r It is the distance the rear vehicle travels during the process of the two vehicles becoming the same speed, and D is the initial distance between the two vehicles.

[0062] When the driving style of the vehicle in front is balanced, τ p (t) represents the current headway; τ is the headway when the driving style of the vehicle in front is identified as conservative. b (t) takes min{0.8τ} p (t), τ min}, where τ min It is the minimum τ that satisfies the constraints in equation (18), and when the driving style of the preceding vehicle is aggressive, τ j (t) is taken as 1.1τ p (t) By adjusting the following distance based on the driving style, the following distance can be shortened when the vehicle in front is identified as conservative, thereby improving driving efficiency and increasing road utilization; when the vehicle in front is identified as aggressive, the following strategy pays more attention to safety and increases the following distance to improve safety.

[0063] 1) Establish a following control strategy model, using the variable distance following model as the upper-level module and the following control strategy module as the lower-level module. The upper-level module outputs the target following distance to the lower-level module, and the following control strategy module implements the input target following distance. This patent uses the LQR optimal control method to track the desired following distance.

[0064] Multi-objective optimization is performed using the relative distance Δd between the foreve vehicle and the vehicle itself, the relative velocity Δv, and the acceleration a of the vehicle itself as state variables. The velocity a of the foreve vehicle is used as the state variable. f Treating it as a disturbance, the state equation can be obtained as follows:

[0065]

[0066] y = Cx

[0067] in u = a; v = a f

[0068] To achieve proper following distance d, the actual distance d should be close to the desired following distance da. e Therefore, Δd should approach 0, and Δv should also approach 0. Regarding the requirement for a, to ensure system comfort, a should be minimized while satisfying the above conditions, i.e., min{|a...} f |}.

[0069] In summary, the control index function is established as follows:

[0070]

[0071] In the formula, q1 and q1 are the weighting coefficients of the control error, and r is the weighting coefficient of the control input. By minimizing the control index function J, the minimum expected following acceleration that satisfies the optimal control of the system can be obtained.

[0072] The positive effect of this invention is that by recognizing the driving style of the vehicle in front and adjusting the following strategy of the adaptive cruise system, driving efficiency and road utilization can be improved while ensuring safety, which plays a positive role in improving traffic conditions and saving energy. Attached Figure Description

[0073] Figure 1 This is an overall flowchart of the technical route of the present invention.

[0074] Figure 2 This is a histogram of the velocity distribution of the sample.

[0075] Figure 3 This is a histogram showing the distribution of the absolute values ​​of acceleration for the sample.

[0076] Figure 4 This is a histogram of the velocity distribution based on the standard deviation of the absolute values ​​of acceleration for the sample.

[0077] Figure 5 This shows the headway distribution for all aggressive vehicles.

[0078] Figure 6 This shows the headway distribution for all conservative vehicles.

[0079] Figure 7 This is a three-dimensional plot of the clustering results under the three-class classification of three-dimensional clustering feature variables. Detailed Implementation

[0080] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 As shown, an improved ACC control method based on the driving style recognition of the preceding vehicle is characterized by the following specific steps:

[0081] Step 1: Select a vehicle trajectory dataset (e.g., the NGSIM vehicle trajectory dataset) as training data and preprocess the data. This step mainly involves the selection of feature vectors, which is based on the adaptive cruise control of the preceding vehicle's driving style. Currently, research on driving style classification mainly focuses on the driving style of the driver of the vehicle itself. Compared to the vehicle itself, there is less data available for the preceding vehicle. The data of the preceding vehicle mainly consists of speed and acceleration. The mean and variance of speed and acceleration are used to reflect the power input of the preceding vehicle's driver when pressing the brake pedal, thereby identifying the driving style of the preceding vehicle's driver.

[0082] Step two involves classifying the processed sample data. In this process, the fuzzy clustering (FMC) method is used for style classification. Compared to the hard clustering methods, primarily K-means clustering, which are mainly used in existing research, the advantage of soft clustering is that the membership degree of each cluster is not simply either "belongs" or "does not belong," but can simultaneously belong to multiple clusters. The membership degree of a sample point to each cluster can be represented by its membership degree, ranging from 0 to 1. This fuzzy clustering is more suitable for applications involving driving style classification. Since the transition from conservative to aggressive driving style lacks a clear dividing line and is relatively gradual, the FMC clustering method can effectively improve the rationality of driving style classification.

[0083] Step three involves establishing an evaluation index for the clustering results to select the optimal clustering method. This index determines the dimension of the feature vector and the number of clusters that yield the best results. Traditional external evaluation indexes require correct labels, making them unsuitable for analyzing clusters lacking proper external labels, such as driving style. By using headway distance as a relatively accurate evaluation index, the improved Land coefficient (ARI) of the clustering results is calculated. This reflects the correlation between the classification results and headway distance. Furthermore, by plotting the headway distance distribution, the accuracy of classification under different methods can be compared, thus achieving the goal of evaluating the clustering effect and selecting the optimal clustering method.

[0084] Step four involves establishing a following model for adaptive cruise control. This model consists of two parts: a variable-distance following model and a following control strategy model. The following model serves as the upper-level model to output the desired following distance, while the control strategy model acts as the lower-level model to achieve this desired following distance. The upper-level model outputs different following distances based on the driving style of the preceding vehicle. Typically, safety-oriented adaptive cruise control systems tend to have a conservative following distance, which may reduce driving efficiency and road utilization. Adaptive cruise control systems based on the preceding vehicle's driving style, when detecting a conservative driving style, will appropriately reduce the following distance while meeting safety constraints to improve driving efficiency and road utilization.

[0085] The specific implementation steps are as follows:

[0086] Step 1: Select the vehicle trajectory dataset as training data and preprocess the data:

[0087] 3) Slice the data and extract the segments from each vehicle's data that do not involve lane changes or steering. Perform isochronous processing on the data, selecting the first 100 frames of motion for each data set and extracting the mean velocity, mean acceleration, and variance as feature values.

[0088] 4) To avoid the influence of different units of measurement of various evaluation indicators, the data is normalized. First, a bar chart is drawn to show the distribution of the extracted data from the dataset, such as... Figures 2-4 As shown, observation of the data reveals that a large amount of data is concentrated in a small subset of the entire value range, with a few outliers influencing the maximum and minimum values ​​of the samples. Therefore, Z-Score standardization is used to preprocess the data to avoid the influence of outliers and to data-normalize the data to remove the influence of dimensions. The Z-Score formula is as follows:

[0089] z=(x-μ) / σ (1)

[0090] Where x is the original score, μ is the mean score of the overall sample space, and σ is the standard deviation of the overall sample space.

[0091] Step two: Fuzzy C-means clustering (FMC) is used to classify the processed sample data, specifically as follows: First, an m-dimensional evaluation index is selected, and feature vectors X1, X2, ..., X are obtained from n sets of preprocessed data. n The sample dataset is then D = {X1, X2, ..., X...} n}, where sample X i ={x1,x2,...,x m Let} be an m-dimensional vector. Given a predetermined number of categories c, then C1, C2, ..., C... CGiven c clusters obtained by clustering dataset D, u i (i = 1, 2, ..., c) are the cluster centers of the i-th cluster, and a single sample X in D is given. k For the i-th cluster C i The membership degree is α ik (i=1,2,...,c;k=1,2,...,n),α ik The membership matrix U that constitutes c*n must satisfy the following conditions:

[0092] α ik ∈[0,1](1≤i≤c,1≤k≤n) (2)

[0093]

[0094] The following weighted distance is used to reflect the metric sample X. k and the i-th cluster C i The correlation between them is expressed by the following formula:

[0095]

[0096] Where u ij For the i-th cluster C i The j-th component of the cluster center, where m is the number of components of the feature vector.

[0097] Therefore, the sum of the distances of all samples to all clusters is:

[0098]

[0099] To adjust the influence of membership degree on distance calculation, the above formula is changed to a weighted sum of distances:

[0100]

[0101] In the formula, p > 1 is a fuzzy parameter, which determines the fuzziness of the clustering, that is, the degree to which a sample point can belong to multiple clusters. The optimal solution is when the weighted distance reaches its minimum value, and Equation (2) is used as a constraint condition for the membership degree, transforming the problem into a constrained minimum value problem. Therefore, the Lagrange multiplier method is used to solve the above problem, and the Lagrange function is:

[0102]

[0103] Where λ1, λ2, ..., λ k It is a Lagrange operator, x kj It is the j-th variable of the k-th sample, u ij It is the j-th variable of the i-th cluster center.

[0104] Let F(α) ik ,λ1,λ2,...,λ k For each α ik , λ k Taking the derivative, we have:

[0105]

[0106]

[0107] From equation (9), we can obtain:

[0108]

[0109] Substitute it And by organizing the information, we can finally arrive at the following solution:

[0110]

[0111] To find the optimal cluster centers, the objective function F(α) is used. ik ,λ1,λ2,...,λ k ) is regarded as the cluster center u ij The function, i.e., F(u) ij By taking its derivative, the optimal cluster center coordinates u of each cluster can be calculated. i as follows:

[0112]

[0113] The formula for calculating the optimal cluster centers is as follows:

[0114]

[0115] Therefore, an iterative method can be used to cluster the data and obtain the optimal clustering result, namely the optimal membership matrix U and the optimal cluster centers u. j The specific steps are as follows:

[0116] 4) Determine the number of clusters, determine the initial parameter p, specify the allowable relative error, i.e., U(t)-U(t-1)<ζ, or specify the maximum number of iterations T;

[0117] 5) Randomly select the initial cluster centers u (0) Calculate the initial membership degree α using the above formula. (0) ;

[0118] 6) Calculate u again (1) ,α (1) The iteration stops when the relative error is less than the allowable error ζ or the number of iterations reaches the specified number.

[0119] Step 3: Establish evaluation metrics for clustering results to determine the optimal dimension of the feature vector and the number of clusters.

[0120] Because it is impossible to obtain an absolutely accurate classification of driving styles, traditional external evaluation metrics cannot be used for evaluation.

[0121] Using improved external metrics, we selected three-dimensional feature vectors (mean velocity, mean absolute acceleration, and standard deviation of acceleration) and two-dimensional feature vectors (mean absolute acceleration and standard deviation of acceleration) to perform binary and three-dimensional clustering, respectively, and compared the clustering effects of a total of four clustering methods.

[0122] Considering that headway can reflect driving style to some extent, and that headway of the vehicle in front is unavailable in real-world conditions and therefore cannot be used as a feature vector for clustering, headway is used as a label for external evaluation. Regarding safe headway standards, the European Conference of Road Directors (CEDR) completed an assessment report on driving safety distance rules and practices in EU member states in 2010. The report concluded that many countries use the "2-second rule" as an important rule of thumb, incorporating it into driving school training and using relevant traffic signs and markings on the road. Furthermore, European Directive 2004 / 54 stipulates that, under normal circumstances, small passenger cars should maintain a distance of at least 2 seconds from the vehicle in front in tunnel sections within the EU. my country's regulations on safe driving distance do not use the headway as a standard. However, the public safety recommended standard "Safe and Civilized Operation Specifications for Motor Vehicle Drivers" (hereinafter referred to as "Operation Specifications") (GA / T 1773.1—2021) issued in 2021 stipulates that the longitudinal safe distance between a small car and the vehicle in front should be greater than the distance the vehicle travels in 3 seconds.

[0123] Therefore, according to domestic and international regulations regarding following distance during vehicle following, vehicles with a following distance of 0-2 seconds are classified as aggressive, because this range of following distance is considered illegal or unsafe in many EU member states and countries such as China. This patent uses following distance as the basis for category labeling and employs two evaluation methods:

[0124] 1) Evaluation was performed using the improved adjusted Rand coefficient, with the following specific steps:

[0125] First, calculate the Rand coefficient RI:

[0126]

[0127] Here, TP represents the number of data points that belong to the same category in both external and internal metrics; TN represents the number of data points that do not belong to the same category in either external or internal metrics; FP represents the number of data points that do not belong to the same category in external metrics but belong to the same category in internal metrics; and FN represents the number of data points that belong to the same category in external metrics but do not belong to the same category in internal metrics. The four types of data point pairs in the denominator encompass all possible outcomes for any two randomly selected data points in the entire sample, thus simplifying the formula to:

[0128]

[0129] Clearly, TP and TN are correct decisions, while FP and FN are incorrect decisions. The RI value ranges from [0,1]. A drawback of RI is that it cannot make the value close to 0 in the case of completely random clustering. Therefore, the adjusted Rand coefficient (ARI) is used. The formula for calculating ARI from RI is as follows:

[0130]

[0131] 2) Visualize and compare using the headway distribution map method. The specific steps are as follows:

[0132] Since the average vehicle headway distance for each driving style category obtained through clustering is distributed across various intervals, the proportion of different average vehicle headway distances for each style under each clustering method is extracted and plotted in a bar chart, thus creating the vehicle headway distance distribution map for each style. Then, several sets of vehicle headway distance distribution maps for the same style under different clustering conditions are plotted on a single graph for comparison. Figures 5-6 As shown, the horizontal axis (1-4) represents the four time intervals from low to high, and the vertical axis represents the data proportion of each interval, which can intuitively compare the quality of clustering.

[0133] For the conservative style clustering, the average headway is used as the criterion. There should be more vehicles distributed in the high headway portion. Therefore, the more accurate headway distribution of the conservative clustering should be mainly distributed in the part with higher headway.

[0134] For the more aggressive type, the more accurate aggressive clustering method should mainly focus on the part with lower headway distance.

[0135] The height of the vehicle's front end is determined by the 2-second limit mentioned above, which was obtained by consulting regulations.

[0136] This invention selects a three-dimensional, three-class classification clustering method. The clustering results are as follows: Figure 7

[0137] Step 4: Establish a following model for adaptive cruise control. This model is characterized by adjusting the following distance based on the driving style of the vehicle in front, as detailed below:

[0138] 2) First, establish a variable-distance following model:

[0139] d e =τ×v+d0 (17)

[0140] In the formula, d e For the desired following distance, τ is a coefficient related to the headway, which can be expressed as τ(t), v is the vehicle speed, and d0 is the minimum following distance. The safe following distance is calculated using the variable headway method, that is, τ is changed according to different driving conditions to adjust the following distance. The adjustment of τ is based on the driving style classification mentioned above and must meet the following constraints:

[0141] d f -d r +D≥d e (18)

[0142] Where d f d is the distance the car in front travels during the process of both cars reaching the same speed. r It is the distance the rear vehicle travels during the process of the two vehicles becoming the same speed, and D is the initial distance between the two vehicles.

[0143] When the driving style of the vehicle in front is balanced, τ p (t) represents the current headway; τ is the headway when the driving style of the vehicle in front is identified as conservative. b (t) takes min{0.8τ} p (t), τ min}, where τ min It is the minimum τ that satisfies the constraints in equation (18), and when the driving style of the preceding vehicle is aggressive, τ j (t) is taken as 1.1τ p (t) By adjusting the following distance based on the driving style, the following distance can be shortened when the vehicle in front is identified as conservative, thereby improving driving efficiency and increasing road utilization; when the vehicle in front is identified as aggressive, the following strategy pays more attention to safety and increases the following distance to improve safety.

[0144] 5) Establish a following control strategy model, using the variable distance following model as the upper-level module and the following control strategy module as the lower-level module. The upper-level module outputs the target following distance to the lower-level module, and the following control strategy module implements the input target following distance. This patent uses the LQR optimal control method to track the desired following distance.

[0145] Multi-objective optimization is performed using the relative distance Δd between the foreve vehicle and the vehicle itself, the relative velocity Δv, and the acceleration a of the vehicle itself as state variables. The velocity a of the foreve vehicle is used as the state variable. f Treating it as a disturbance, the state equation can be obtained as follows:

[0146]

[0147] y = Cx

[0148] in u = a; v = a f

[0149] To achieve proper following distance d, the actual distance d should be close to the desired following distance da. e Therefore, Δd should approach 0, and Δv should also approach 0. Regarding the requirement for a, to ensure system comfort, a should be minimized while satisfying the above conditions, i.e., min{|a...} f |}.

[0150] In summary, the control index function is established as follows:

[0151]

[0152] In the formula, q1 and q1 are the weighting coefficients of the control error, and r is the weighting coefficient of the control input. By minimizing the control index function J, the minimum expected following acceleration that satisfies the optimal control of the system can be obtained.

Claims

1. An improved ACC control algorithm based on the driving style recognition of the preceding vehicle, characterized in that... The specific implementation steps are as follows: Step 1: Select the vehicle trajectory dataset as training data and preprocess the data: 1) Slice the data and extract the segments of each vehicle's data that do not involve lane changing or turning. Perform isochronous processing on the data. Select the first 100 frames of motion for each data set and extract the mean of the velocity, the mean of the acceleration, and the variance as feature quantities. 2) To avoid the influence of different units of measurement of various evaluation indicators, the data is normalized. First, a bar chart is plotted to show the distribution of the extracted data from the dataset. A large amount of data is concentrated in a small subset of the entire value range, and a few outliers affect the maximum and minimum values ​​of the samples. Therefore, Z-Score standardization is used to preprocess the data to avoid the influence of outliers and to normalize the data to remove the influence of units of measurement. The Z-Score formula is as follows: （1） Where x is the original score, and μ is the mean score of the overall sample space. The standard deviation of the population sample space; Step two: Use FMC fuzzy C-means clustering to classify the processed sample data, as follows: First, an m-dimensional evaluation index is selected, and feature vectors X1, X2, ..., X3 are obtained from n sets of preprocessed data. n The sample dataset is then D = {X1, X2, ..., X...} n }, where sample X i ={x1,x2,...,x m Let} be an m-dimensional vector; given a predetermined number of categories c, then C 1， C2, ..., C C For dataset D, c clusters are obtained by clustering. (i=1,2,...,c) are the cluster centers of the i-th cluster, and a single sample X in D is given. k For the i-th cluster C i The membership degree is (i=1,2,...,c;k=1,2,...,n), The membership matrix U that constitutes c*n must satisfy the following conditions: The following weighted distance is used to reflect the metric sample X. k and the i-th cluster C i The correlation between them is expressed by the following formula: in For the i-th cluster C i The j-th component of the cluster center, where m is the number of components of the feature vector; Therefore, the sum of the distances of all samples to all clusters is: To adjust the impact of membership degree on distance calculation, the above formula is changed to a weighted sum of distances: In the formula, p>1 is a fuzzy parameter, which determines the fuzziness of the clustering, that is, the degree to which a sample point belongs to multiple clusters; the optimal solution is when the weighted distance reaches its minimum value, and the formula (2) is used as a constraint condition for the membership degree, transforming the problem into a constrained minimum value problem. Therefore, the Lagrange multiplier method is used to solve the above problem, and the Lagrange function is: in, It is a Lagrange operator. It is the j-th variable of the k-th sample. It is the j-th variable of the i-th cluster center; make For each one Taking the derivative, we have: From equation (9), we get: Substitute it And after organizing the results, we finally obtained: To find the optimal cluster centers, the objective function is... Considered as cluster center The function, i.e. The optimal cluster center coordinates of each cluster are calculated by taking its derivative. as follows: The formula for calculating the optimal cluster centers is as follows: Therefore, by using an iterative method to cluster the data and obtain the optimal clustering result, we can obtain the optimal membership matrix U and the optimal cluster centers. The specific steps are as follows: 1) Determine the number of clusters, determine the initial parameter p, and specify the allowable relative error, i.e. Alternatively, a maximum number of iterations T can be specified; 2) Randomly select the initial cluster centers u （0） Calculate the initial membership degree using the above formula. ; 3) Calculate u again （1） Until the relative error is less than the allowable error Alternatively, iteration may stop when the specified number of iterations is reached. Step 3: Establish evaluation metrics for clustering results to determine the optimal dimension of the feature vector and the number of clusters. Because it is impossible to obtain an absolutely accurate classification of driving styles, it is impossible to use traditional external evaluation indicators for evaluation; Using improved external metrics, three-dimensional eigenvectors (mean velocity, mean absolute acceleration, and standard deviation of acceleration) and two-dimensional eigenvectors (mean absolute acceleration and standard deviation of acceleration) were selected for binary and ternary clustering, respectively. The clustering effects of a total of four clustering methods were compared. Using the headway as the criterion for category labeling, two evaluation methods are employed: 1) Evaluation was performed using the improved adjusted Rand coefficient, with the following specific steps: First, calculate the Rand coefficient RI: Where TP is the number of data points that belong to the same category in both external and internal indicators; TN is the number of data points that do not belong to the same category in either external or internal indicators; FP is the number of data points that do not belong to the same category in external indicators but belong to the same category in internal indicators; and FN is the number of data points that belong to the same category in external indicators but do not belong to the same category in internal indicators. The four types of data point pairs in the denominator encompass all possible outcomes for any two data points selected from the entire sample, thus simplifying the formula to: Clearly, TP and TN are correct decisions, while FP and FN are incorrect decisions. RI ranges from [0,1]. A drawback of RI is that it cannot make the value close to 0 in the case of completely random clustering. Therefore, the adjusted Land coefficient (ARI) is used. The formula for calculating ARI from RI is as follows: 2) Visualize and compare using the headway distribution map method. The specific steps are as follows: Since the average vehicle headway distance of each driving style obtained by clustering is distributed in each interval, the proportion of different average vehicle headway distances in each style under each clustering method is extracted and a bar chart is drawn, which is the vehicle headway distance distribution map of each style. Then, several sets of vehicle headway distance distribution maps of the same style under different clustering conditions are drawn into one map for comparison. For the conservative style clustering, the average headway is used as the criterion. There should be more vehicles distributed in the high headway portion. Therefore, the more accurate headway distribution of the conservative cluster should be mainly distributed in the part with higher headway. For the more aggressive type, the more accurate aggressive clustering method should mainly focus on the part with the lower train headway. The height of the front of the vehicle is divided according to the 2-second limit obtained from the regulations mentioned above; Select a three-dimensional, three-class clustering method; Step 4: Establish a following model for adaptive cruise control. This model is characterized by adjusting the following distance based on the driving style of the vehicle in front, as detailed below: 1) First, establish a variable-distance following model: In the formula, For the desired vehicle distance, A coefficient related to the headway is expressed as: , For the speed of the vehicle, To minimize the required following distance, the safe following distance is calculated using the variable time-distance method, which adjusts the distance according to different driving conditions. This allows for adjusting the following distance; for The adjustments are based on the driving style classifications described above and must meet the following constraints: in It is the distance the car in front travels during the process of both cars reaching the same speed. It is the distance the rear vehicle travels during the process of the two vehicles becoming the same speed, and D is the initial distance between the two vehicles; When the driver of the vehicle in front has a balanced driving style This is the current headway; assuming the driving style of the vehicle in front is identified as conservative, Take min{ , },in, It is the minimum that satisfies the constraints in equation (18). When the driver of the car in front has an aggressive driving style, Pick By adjusting the following distance based on driving style, the system shortens the following distance when the vehicle in front is identified as conservative, thereby improving driving efficiency and increasing road utilization. When the vehicle in front is identified as aggressive, the following strategy prioritizes safety and increases the following distance to improve safety. 3) Establish a following control strategy model, using the variable distance following model as the upper-level module and the following control strategy module as the lower-level module. The upper-level module outputs the target following distance to the lower-level module, and the following control strategy module implements the input target following distance. The LQR optimal control algorithm is used to track the desired following distance. The relative distance between the car in front and the car itself relative speed Vehicle acceleration Multi-objective optimization is performed using the state variable; the speed of the vehicle in front is used as the variable. Treating it as a disturbance, the state equation is as follows: Among them ; ; ; ;u=a; To achieve proper following distance, the actual distance *d* should approximate the desired following distance. It should approach 0. It should also approach 0. Regarding the requirement for 'a', to ensure system comfort, 'a' should be minimized while satisfying the above conditions, i.e. ; In summary, the control index function is established as follows: In the formula , , where are the weighting coefficients for the control error and r is the weighting coefficient for the control input. By minimizing the control index function J, we obtain the minimum expected following acceleration that satisfies the optimal control of the system.

Images