A driving style classification method based on image recognition and topsis comprehensive evaluation

Abstract

The present application relates to a kind of driving style classification method based on image recognition and TOPSIS comprehensive evaluation, comprising the following steps: step 1, the collection of driver driving data, driving data is divided into operation event and routine driving, and further each operation event is divided into multiple sub-events;Step 2, according to operation event duration setting driving style, and using Kmeans one-dimensional clustering algorithm further classifies operation event duration;Step 3, by model obtains the driving style of each sub-event in operation event duration and each speed limit level in routine driving;Step 4, based on the driving style of sub-event confirms the driving style of corresponding operation event.The present application innovatively divides the entire driving data into different events for discussion, and uses different driving style recognition methods for different events, and finally completes classification by the weight fusion of TOPSIS comprehensive score method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving, and particularly relates to a driving style classification method based on image recognition and TOPSIS comprehensive evaluation. BACKGROUND

[0002] With the development of the automobile industry and artificial intelligence, the status of automobiles in people's lives is becoming higher and higher, and the research on automatic driving and intelligent networked vehicles is also increasing. According to the automatic driving level classification standard proposed by the American Automobile Industry Association (SAE), in the development process of the automation degree of automobiles from L0 (No Driving Automation) level to L5 (Full Driving Automation) level, due to the limitations of current technology development, the basic L2 (Partial Driving Automation) level automatic driving vehicles can be commercially used on the market at present, and therefore, the current research is mainly directed to L3 (Conditional Driving Automation) level driving. In the research on L3 level driving, most researchers believe that the research on the allocation of driving rights is a feasible direction.

[0003] In the problem of driving right allocation, the research on driving style classification method is crucial. The classification of driving style can be used as the input of the driving right allocation model to optimize and correct the driving right allocation model, so that the driving right allocation model is more intelligent and humanized. For driving style classification, early research mainly uses subjective methods such as questionnaire surveys to score and classify by experts. With the development of computer network technology, most current research is basically based on the analysis of vehicle data (the most common of which is acceleration data) to classify. Some researchers use machine learning methods such as convolutional neural networks to process and classify data, but in previous research, the routine driving and operation events in the driving process are not distinguished and discussed.

[0004] Previous research and patents usually use two kinds of data sources for driving style, namely network public data sets and vehicle and driver data sets. In terms of data processing and classification methods, there are traditional statistical analysis methods and end-to-end classification methods using machine learning. In terms of driving style classification, some researchers believe that it can be divided into aggressive and normal, and other researchers believe that it should be divided into aggressive, normal, and cautious, and some researchers believe that it should be divided into four driving styles, or more detailed into six driving styles. On the other hand, some researchers divide the driving style into a non-level range between -1 and +1, so as to form a more detailed classification style.

[0005] However, too many classification categories not only bring the increase of computational complexity, but also are not conducive to the simplification and efficiency of the model, while too few classification categories of 2 styles are easy to cause data loss and large error. SUMMARY

[0006] The application provides a driving style classification method based on image recognition and TOPSIS comprehensive evaluation, which comprises the following steps:

[0007] Step 1, collecting the driving data of the driver, dividing the driving data into operation events and regular driving, and further dividing each operation event into multiple sub-events; the sub-events include the entire event duration, the accelerator pedal speed, the accelerator pedal acceleration, the steering wheel rotation speed and the steering wheel rotation acceleration;

[0008] Step 2, setting the driving style according to the operation event duration, and further classifying the operation event duration using a one-dimensional Kmeans clustering algorithm;

[0009] Step 3, obtaining the driving style of each sub-event of the operation event duration and each speed limit level of the regular driving through a model;

[0010] Step 4, confirming the driving style of the corresponding operation event based on the driving style under the sub-event.

[0011] Further, in step 1, the following steps are further included:

[0012] Step 1.1, classifying the regular driving section and the operation event section in the entire driving process by data analysis and statistics; wherein the operation event is further divided into start-stop events, overtaking events and meeting events which are extremely common in the driving process; the regular event is defined as that the driver does not operate the accelerator or the pedal during the driving process, and is specifically divided into the speed change of the automobile under the three speed limit modes of 0-60km / h, 60-80km / h and 80-120km / h on a speed limit road;

[0013] Step 1.2, further subdividing the start-stop events, overtaking events and meeting events under the operation event, and specifically dividing them into five sub-events of the entire event duration, the accelerator pedal speed, the accelerator pedal acceleration, the steering wheel rotation speed and the steering wheel rotation acceleration.

[0014] Further, in step 2, the following steps are further included:

[0015] Step 2.1, dividing the data into three categories according to the three driving styles of aggressive, normal and cautious, and determining the sample center k value of the clustering algorithm as 3;

[0016] Step 2.2, randomly select 3 points from the time data points that last throughout the event duration as the initial sample center of 3 clusters;

[0017] Step 2.3, calculate the distance between the data points and each sample center, and take the cluster where the nearest center is as the cluster to which the data point belongs;

[0018] Step 2.4, calculate the average value of the data points of each cluster, and take it as the new sample center;

[0019] Step 2.5, repeat steps 2.2-2.4 until the sample centers of all clusters no longer change, or the maximum number of iterations is reached, to obtain the driving style on the sub-event of the entire event duration.

[0020] Further, in step 3, the following steps are further included:

[0021] Step 3.1, according to the collected accelerator pedal and steering wheel position data, v i =(x i+1 -x i ) / t and a i =2(x i+1 -x i ) / t 2 Two formulas calculate the speed and acceleration, and generate four curves of accelerator pedal speed, accelerator pedal acceleration, steering wheel angular velocity and steering wheel angular acceleration corresponding to start-stop events, overtaking events and passing events respectively;

[0022] Step 3.2, according to the height of the area formed by the generated speed and acceleration curves and the horizontal axis, different color depths are given to form more easily recognizable speed and acceleration images.

[0023] Step 3.3, input a large number of different speed and acceleration images into a super-light image classification model for classification and recognition.

[0024] Further, in step 3.3, the following steps are further included:

[0025] Step 3.3.1, for the original picture, through data enhancement strategy integration, three data enhancement methods of image transformation, image cropping and image aliasing are fused, the generalization ability of the model is enhanced by adjusting the trigger probability p%, and the processed picture is obtained in this way. The processed and enhanced picture is taken as the output into the semi-supervised distillation algorithm SSDL for weight pre-training through this model;

[0026] Step 3.3.2, the unlabeled picture data obtained in step 3.3.1 above is learned to the knowledge of ImageNet 22k unlabeled large-scale data through a small model, and the weights obtained through the semi-supervised distillation algorithm SSDL are used as the initialization parameters of the model, and the backbone network model PP-LCNet is entered;

[0027] Step 3.3.3, the weights trained by the algorithm of step 3.3.2 are used to initialize the backbone network model PP-LCNet, and then the picture data is input into the model for optimization. After being processed by the network, a fully connected layer network is obtained, and the result is input into the SKL distillation loss function.

[0028] Step 3.3.4, the picture data in step 3.3.3 is input into the classroom pre-training model for training, and then the SKL-UGI knowledge distillation algorithm is used to introduce symmetric information on the basis of the classic KL knowledge distillation algorithm. The picture output by the algorithm is input into the SKL distillation loss function to obtain the final image recognition result.

[0029] Through the recognition result, the driving style of the driver under the three operating events of start-stop, overtaking and meeting is obtained, as well as the four sub-events of accelerator pedal speed, accelerator pedal acceleration, steering wheel angular velocity and steering wheel angular acceleration, and the driving style under the three speed limit modes of 60km / h, 60-80km / h and 80-120km / h in normal driving.

[0030] Further, in step 4, the following steps are further included:

[0031] Step 4.1, according to the three events of start-stop, overtaking and meeting, the number of samples 3 and the number of sample dimensions 3 of the three driving styles of aggressive, normal and cautious are determined, the driving style under each sub-event is obtained, and a value is determined for each driving style according to the influence size, which is used as the value x of the random variable in the algorithm ij , that is, the variable matrix is obtained:

[0032]

[0033] Wherein, i respectively represents the three events of start-stop, overtaking and meeting, and j represents the three driving styles of aggressive, normal and cautious;

[0034] Step 4.2, the above matrix is standardized, that is, according to the meaning of different indicators, the formula and are used to divide it into positive indicators and negative indicators, wherein x j represents the variable value of the three driving styles under the same event;

[0035] Step 4.3, the formula wherein The entropy value of each dimension in the data is calculated;

[0036] Step 4.4, the redundancy degree is calculated by the formula d j = 1-E j The redundancy degree is calculated by the formula The weight of each redundancy degree is calculated, and the weighted normalized matrix Z=(Z ij ) n×m =(p ij *ω j );

[0037] Step 4.5, the positive ideal solution and the negative ideal solution are determined by the best value and the worst value in the sample respectively, and the distance of each sample to the positive ideal solution and the distance of each sample to the negative ideal solution are calculated according to the formula

[0038] Step 4.6, the closeness of each evaluation object to the optimal scheme is planned by the following formula The result of the closeness is in the range of [0, 1], and the closer to 1, the better the sample score is; the score value is the weight of the driving style under each sub-event, which indicates the influence degree of the driving style on the final result, and the maximum weight value is taken as the performance of the driving style under the event;

[0039] Step 4.7, by the above method, steps 4.1-4.6 are repeated to perform the bottom layer weight fusion on the entire event duration, the accelerator pedal speed, the accelerator pedal acceleration, the steering wheel angular velocity and the steering wheel angular acceleration of the start-stop event, the overtaking event and the meeting event under each of the five sub-events; steps 4.1-4.6 are repeated to perform the secondary top layer weight fusion on the 60km / h, 60-80km / h and 80-120km / h three speed limit modes in the regular driving; and steps 4.1-4.6 are repeated to perform the top layer weight fusion on the operation event and the regular driving, so that the driving style of the driver is finally obtained, and the classification of the driving style is completed.

[0040] The beneficial effects achieved by the present application are:

[0041] Compared with the existing driving style classification methods using only vehicle speed and acceleration or using machine learning, the whole driving data is discussed in different events, different driving style recognition methods are used for different events, and finally the classification is completed by weight fusion through the TOPSIS comprehensive scoring method. The classic Kmeans clustering method is used for classification according to the whole event duration, and higher classification accuracy can be achieved in the case of small amount of data; the super-light image classification (PLUC) model is innovatively used for classification according to the speed and acceleration images of the accelerator pedal and steering wheel angle, which effectively ensures the image classification accuracy in the case of small amount of data; finally, the TOPSIS comprehensive scoring method based on entropy weight method is used for multi-layer weight fusion of the driving styles of each event, so that the classification of driving style is more objective, the classification credibility is improved, the classification algorithm and model are simplified, and the classification of driving style provides better model input parameters for the research of driving right allocation. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 FIG. 1 is a flowchart of a driving style classification method based on image recognition and TOPSIS comprehensive evaluation;

[0043] Figure 2 FIG. 2 is a flowchart of a PLUC algorithm in a driving style classification method based on image recognition and TOPSIS comprehensive evaluation;

[0044] Figure 3 FIG. 3 is a schematic diagram of the speed or acceleration change curve in each event and three speed limit modes in a driving style classification method based on image recognition and TOPSIS comprehensive evaluation;

[0045] Figure 4 FIG. 4 is a weight distribution diagram of three events in the operation event of a driving style classification method based on image recognition and TOPSIS comprehensive evaluation. DETAILED DESCRIPTION

[0046] The technical solutions of the present application will be described in more detail below with reference to the accompanying drawings. The present application includes but is not limited to the following embodiments.

[0047] As shown in the accompanying Figure 1 , the present application proposes a driving style classification method based on image recognition and TOPSIS comprehensive evaluation. The method includes the following steps:

[0048] Step 1: Collect the driving data of the driver through the driving simulator or real vehicle data collection method, and segment the whole driving data according to the change characteristics of the driving vehicle speed and acceleration.

[0049] Further, the specific distinguishing step is:

[0050] Step 1.1, classify the regular driving section and the operation event section in the entire driving process by means of data analysis statistics. The operation event is further divided into start-stop event, overtaking event and meeting event which are extremely common in the process of driving. The regular event is defined as that the driver does not operate the accelerator or the pedal during driving, and is specifically divided into three speed change modes of 0-60km / h, 60-80km / h and 80-120km / h on a speed-limited road.

[0051] Step 1.2, the start-stop event, overtaking event and meeting event under the operation event are further divided, and each of them is divided into five sub-events, including the entire event duration, the accelerator pedal speed, the accelerator pedal acceleration, the steering wheel rotation speed and the steering wheel rotation acceleration.

[0052] Step 2, in the above five sub-events, for the driving style shown in the entire event duration sub-event, considering the simplicity of classification, if the time duration is very long, the driving style of the driver is biased towards the cautious driving style, on the contrary, if the time duration is very short, it is considered to be more biased towards the aggressive driving style, and between the two, it is biased towards the normal driving style. The data of the entire event duration part is classified by using the Kmeans one-dimensional clustering algorithm. Further, the specific steps are:

[0053] Step 2.1, according to the three driving styles of aggressive, normal and cautious defined in the application, the data is divided into three categories, that is, the sample center k value of the clustering algorithm is determined to be 3;

[0054] Step 2.2, randomly select three points from the time data points in the entire event duration as the initial sample centers of the three clusters;

[0055] Step 2.3, calculate the distance between the data points and each sample center, and take the cluster where the nearest center is located as the cluster to which the data point belongs;

[0056] Step 2.4, calculate the average value of the data points of each cluster, and take it as the new sample center;

[0057] Step 2.5, repeat steps 2.2-2.4 until the sample centers of all clusters no longer change, or the maximum iteration number is reached, to obtain the driving style on the entire event duration sub-event.

[0058] Step 3, for the four sub-events of the accelerator pedal speed, accelerator pedal acceleration, steering wheel angle speed and steering wheel angle acceleration in the above operation events, and the identification of driving style in the three speed limit modes of 60-80 km / h and 80-120 km / h in the conventional driving. The curves of speed or acceleration with time in each sub-event and three speed limit modes, the curves of speed with time in three speed limit modes; the horizontal axis of the curve is time, and the vertical axis is speed value, and different color depths (from black (R=0 G=0 B=0) to white (R=255 G=255 B=255)) are given according to the height of the speed value at each time scale, so as to form a one-dimensional image as shown in the accompanying Figure 3 .

[0059] According to the one-dimensional image generated by the data, the classification is carried out by feature extraction and image recognition, and further, the steps are:

[0060] Step 3.1, according to the collected accelerator pedal and steering wheel position data, v i =(x i+1 -x i ) / t and a i =2(x i+1 -x i ) / t 2 Two formulas are used to calculate the speed and acceleration, and four curves of accelerator pedal speed, accelerator pedal acceleration, steering wheel angle speed and steering wheel angle acceleration corresponding to start-stop event, overtaking event and meeting event are generated;

[0061] Step 3.2, according to the area formed by the generated speed and acceleration curves and the horizontal axis, different color depths are given according to the height of the vertical coordinate, and a speed and acceleration image that is easier to identify is formed.

[0062] Step 3.3, input a large number of different speed and acceleration images into a super-light image classification model (PLUC) for classification and identification. As shown in the accompanying Figure 2 , the steps are further shown as:

[0063] Step 3.3.1, for the original picture, through data enhancement strategy integration, three data enhancement methods of image transformation, image cropping and image aliasing are fused, the generalization ability of the model is enhanced by adjusting the trigger probability p%, and the picture after processing is obtained. The processed and enhanced pictures are taken as output by the model and enter the semi-supervised distillation algorithm SSDL for weight pre-training;

[0064] Step 3.3.2, the unlabeled picture data obtained in step 3.3.1 above is learned to the knowledge of ImageNet 22k unlabeled large-scale data by a small model, and the weights obtained by the semi-supervised distillation algorithm SSDL are used as the initialization parameters of the model, and then the backbone network model PP-LCNet is entered;

[0065] Step 3.3.3, the weights trained by the algorithm of step 3.3.2 are used to initialize the backbone network model PP-LCNet, and then the picture data (including labeled and unlabeled data) are input into the model for optimization. After the network processing, a fully connected layer network (FC) is obtained, and the result is input into the SKL distillation loss function.

[0066] Step 3.3.4, the picture data in step 3.3.3 is input into the classroom pre-training model for training, and then the SKL-UGI knowledge distillation algorithm is used to introduce symmetric information on the basis of the classical KL knowledge distillation algorithm. The picture output by the algorithm is input into the SKL distillation loss function to obtain the final image recognition result. Through the recognition result, the driving style of the accelerator pedal speed, the accelerator pedal acceleration, the steering wheel angular velocity and the steering wheel angular acceleration under the three operation events of start-stop, overtaking and meeting, as well as the driving style under the three speed limit modes of 60km / h, 60-80km / h and 80-120km / h in the normal driving are obtained.

[0067] Step 4, based on the driving style obtained under each sub-event, if the driving style classification results obtained under each sub-event are the same, the driving style under the event can be determined; if the driving style classification results obtained under each sub-event are different, it is regarded as an abnormal value and is judged by using the weight fusion method. Since there are five sub-events under each event, the five-dimensional space of the operation event is not conducive to visualization, so the influence of the three-dimensional space of the style classification composed of the overtaking event, the start-stop event and the meeting event on the operation event is taken as an example, as shown in Figure 4As shown, the dark gray blocks between each pair of events are normal values of the same driving style judgment result (i.e., judged as aggressive, normal or cautious between each pair of events); the light gray blocks are adjacent abnormal values of the driving style judgment result (i.e., judged as aggressive and normal, or judged as normal and cautious between each pair of events); the black blocks are opposite abnormal values of the driving style judgment result (i.e., judged as aggressive and cautious between each pair of events); the black arrow area is the normal value area of the judgment style between each pair of events; and the white arrow area (i.e., the diagonal area of the three-dimensional space in the figure) is the normal value area of the judgment style between all events. If the same driving style (i.e., normal value of aggressive, normal or cautious) is exhibited under three events (i.e., the same driving style is exhibited under the operation events), it can be considered that the driving style is exhibited under the operation events; if different driving styles are exhibited under three events (i.e., abnormal values in the non-diagonal area of the three-dimensional space in the figure), the TOPSIS comprehensive score based on the entropy weight method is used for weight fusion of the next top layer (i.e., the three event layers of starting and stopping, overtaking and meeting under the operation events, and the three speed limit mode layers of 60 km / h, 60-80 km / h and 80-120 km / h under the regular driving in the future). The specific steps are as follows:

[0068] Step 4.1, according to the three events of starting and stopping, overtaking and meeting, the sample number 3 and the sample dimension number 3 of the three driving styles of aggressive, normal and cautious are determined, the driving style under each sub-event is obtained, and a value is given to each driving style according to the influence size, which is the value of the random variable x in the algorithm ij , that is, the variable matrix

[0069]

[0070] wherein i represents the three events of starting and stopping, overtaking and meeting, and j represents the three driving styles of aggressive, normal and cautious;

[0071] Step 4.2, the above matrix is standardized, that is, according to the meaning of different indexes, the formula and are used to divide them into positive indexes and negative indexes, wherein x j represents the variable value of the three driving styles under the same event;

[0072] Step 4.3, the entropy value of each dimension in the data is calculated by the formula (wherein );

[0073] Step 4.4, the redundancy is calculated by the formula d j = 1-E j , and the formula The weight calculation of each redundancy is performed, and the weighted normalized matrix Z = (Z ij ) n×m = (p jj * ω j );

[0074] Step 4.5, the positive ideal solution and the negative ideal solution are determined by the best value and the worst value in the sample respectively. The distance of each sample to the positive ideal solution and the distance of each sample to the negative ideal solution are calculated according to the formula

[0075] Step 4.6, the closeness of each evaluation object to the optimal scheme is planned according to the following formula The result has a value range of [0, 1], and the closer to 1, the better the sample score. The score value is the weight of the driving style under each sub-event, indicating the influence degree on the final result, and the maximum weight value is taken as the driving style performance under the event.

[0076] Step 4.7, through the above method, if the driving style of the entire event duration, accelerator pedal speed, accelerator pedal acceleration, steering wheel angular velocity and steering wheel angular acceleration of the start-stop event, overtaking event and passing event are all judged to be the same (i.e. all are aggressive, normal or cautious) by the above clustering and image recognition method, then the driving style of the start-stop, overtaking or passing event is obtained; if the styles judged by the five sub-events are not all the same, then steps 4.1-4.6 are repeated for bottom layer (i.e. the entire event duration, accelerator pedal speed, accelerator pedal acceleration, steering wheel angular velocity and steering wheel angular acceleration of the start-stop, overtaking and passing events) weight fusion, the difference is that the number of samples and the dimension in step 4.1 are both 5 for the five sub-events, and the variable matrix obtained is

[0077]

[0078] Step 4.8, similarly, if the driving styles classified by the speed curve changed into one-dimensional image and then using image recognition method are the same in the three speed limit modes of 60km / h, 60-80km / h and 80-120km / h in regular driving, then the driving style performed in regular driving can be determined; if not, then steps 4.1-4.6 are repeated for secondary top layer weight fusion of the driving styles judged in the three modes;

[0079] Step 4.9, after the style of the operation event is determined through steps 4.1-4.6 and the style of the performance under the regular driving is determined through step 4.8, if the styles determined by both are the same (all are aggressive, normal or cautious), the driving style of the driver is obtained, and the classification of the driving style is completed; if the styles determined by both are different, steps 4.1-4.6 are repeated to perform the top layer (i.e. the operation event and the regular driving layer) weight fusion of the driving styles of the operation event and the performance under the regular driving, the difference is that, since there are only two indexes of the operation event and the regular driving, the sample number and the sample dimension number in step 4.1 are both 2, and the obtained variable matrix is Then, 4.2-4.6 are sequentially completed, and finally the driving style of the driver is obtained, and the classification of the driving style is completed.

[0080] The present application is not limited to the above-mentioned specific embodiments, and those skilled in the art can implement the present application by using other various specific embodiments according to the content disclosed in the embodiments and the drawings, therefore, any design using the design structure and the idea of the present application, and making some simple changes or modifications, falls within the protection scope of the present application.

Claims

1. A driving style classification method based on image recognition and TOPSIS comprehensive evaluation, characterized in that: The driving style classification method based on image recognition and TOPSIS comprehensive evaluation includes the following steps: Step 1: collecting driver driving data, dividing the driving data into operational events and regular driving, and further dividing each operational event into multiple sub-events; the sub-events include the entire event duration, accelerator pedal speed, accelerator pedal acceleration, steering wheel angular velocity, and steering wheel angular acceleration; Step 2: Set the driving style according to the duration of the operation event and use the Kmeans one-dimensional clustering algorithm to further classify the duration of the operation event; Step 3: Obtain the driving style of each sub-event in the operation event except the operation event duration and each speed limit level in regular driving through the model; Step 4: confirming the driving style corresponding to the operation event based on the driving style under the sub-event; Step 3 also includes the following steps: Step 3.1, based on the collected accelerator pedal and steering wheel position data, v i =(x i+1 -x i ) / t and a i =2(x i+1 -x i ) / t 2 Two formulas are used to calculate its speed and acceleration, and four curves are generated: accelerator pedal speed, accelerator pedal acceleration, steering wheel angular velocity, and steering wheel angular acceleration, corresponding to start-stop events, overtaking events, and meeting events. Step 3.2: assign different color depths based on the height of the vertical coordinate of the area enclosed by the generated velocity and acceleration curve and the horizontal axis, forming a velocity and acceleration image that is easier to identify; Step 3.3: Input a large number of different speed and acceleration images into the ultra-lightweight image classification model for classification and recognition; In step 4, the bottom-level weight fusion is performed on the five sub-events of the entire event duration, accelerator pedal speed, accelerator pedal acceleration, steering wheel angular velocity, and steering wheel angular acceleration in the start-stop event, overtaking event, and meeting event; the second-level weight fusion is performed on the three speed limit modes of 60km / h, 60-80km / h, and 80-120km / h in regular driving; and the top-level weight fusion is performed on the operation event and regular driving. Finally, the driver's driving style is obtained and the driving style classification is completed.

2. The driving style classification method based on image recognition and TOPSIS comprehensive evaluation according to claim 1 is characterized in that: Step 1 also includes the following steps: Step 1.1: Analyze and statistically analyze data to classify regular driving sections and maneuver event sections throughout the entire driving process. Maneuver event sections are further subdivided into start-stop events, overtaking events, and meeting events, which are extremely common in normal driving. Regular events are defined as the driver not performing corresponding accelerator or pedal operation during driving. Specifically, they are categorized into vehicle speed changes within three speed limit modes: 0-60 km / h, 60-80 km / h, and 80-120 km / h on a specific speed-limited road. In step 1.2, the start-stop event, overtaking event, and meeting event under the operation event are further subdivided into five sub-events each: the entire event duration, accelerator pedal speed, accelerator pedal acceleration, steering wheel angular velocity, and steering wheel angular acceleration.

3. The driving style classification method based on image recognition and TOPSIS comprehensive evaluation according to claim 2 is characterized in that: Step 2 also includes the following steps: Step 2.1: Divide the data into three categories based on the three driving styles of aggressive, normal, and cautious, and determine the sample center k value of the clustering algorithm to be 3; Step 2.2, randomly select 3 points from the time data points of the entire event duration as the initial sample centers of the 3 clusters; Step 2.3: Calculate the distance between the data point and each sample center, and take the cluster with the closest center as the cluster to which the data point belongs; Step 2.4, calculate the mean of the data points of each cluster and use it as the new sample center; Step 2.5: Repeat steps 2.2 to 2.4 until the sample centers of all clusters no longer change or the maximum number of iterations is reached, and the driving style of the entire event duration is obtained.

4. The driving style classification method based on image recognition and TOPSIS comprehensive evaluation according to claim 1 is characterized in that: Step 3.3 also includes the following steps: In step 3.3.1, for the original image, we integrate the data augmentation strategies and fuse three data augmentation methods: image transformation, image cropping, and image aliasing. By adjusting the trigger probability p%, we enhance the generalization ability of the model to obtain the processed image. This model uses the processed and enhanced image as the output and feeds it into the semi-supervised distillation algorithm SSLD for weight pre-training. In step 3.3.2, for the unlabeled image data obtained in step 3.3.1 above, the small model learns the knowledge of ImageNet22k unlabeled large-scale data, and uses the weights obtained by pre-training with the semi-supervised distillation algorithm SSLD as the model initialization parameters to enter the backbone network model PP-LCNet; In step 3.3.3, the backbone network model PP-LCNet is initialized using the weights trained by the algorithm in step 3.3.

2. The image data is then input into the model for optimization. After processing the network, a fully connected layer network is obtained, and the result is input into the SKL distillation loss function. In step 3.3.4, the image data from step 3.3.3 is fed into the classroom pre-trained model for training. The SKL-UGI knowledge distillation algorithm is then used to introduce symmetric information based on the classic KL knowledge distillation algorithm. The image output by the algorithm is fed into the SKL distillation loss function to obtain the final image recognition result. The recognition results provide four sub-events: accelerator pedal speed, accelerator pedal acceleration, steering wheel angular velocity, and steering wheel angular acceleration for the three operational events of start-stop, overtaking, and meeting, as well as the driving style represented by the three speed limit modes of 60 km / h, 60-80 km / h, and 80-120 km / h under normal driving.

Images