Recognition Of Aggressive Driving Behavior Based On Calculation Of Vehicle Mobile Sensing Data

Aggressive driving behavior is one of the main causes of traffic accidents.Accurate recognition of aggressive driving behavior is the premise to timely and effectively conducting warning or intervention to the driver.The research on aggressive driving behavior recognition is of great significance for improving road traffic safety.Existing data-driven aggressive driving behavior recognition methods have disadvantages such as improper processing of the dataset with imbalanced class distribution,high miss rate,and poor recognition accuracy.Aiming to deal with these disadvantages,starting from the features of the aggressive driving behavior dataset,and using deep learning and ensemble learning methods,an aggressive driving behavior recognition method is built in this paper,the main research contents are as follows:（1）Experimental system construction and experimental organization.An integrated experimental vehicle for driving behavior and safety based on the multi-sensor array is constructed.Real vehicle experiments are designed and organized based on the experimental system,and a naturalistic driving dataset containing aggressive driving behavior data is acquired.（2）Feature parameter selection and data processing.Firstly,according to the definition and characteristics of aggressive driving behavior,the feature parameters are selected.Secondly,the data is standardized and processed by the sliding window,and the dataset is divided into the training set and the testing set.Finally,aiming at the problem of imbalanced distribution between aggressive driving behavior sequence samples and normal driving behavior sequence samples,the training set is balanced using the self-organizing map-based dataset balancing method,the two-level clustering-based dataset balancing method,and the evenly divided dataset balancing method.The results show that,compared with the dataset balancing method based on self-organizing map and the dataset balancing method based on two-level clustering,the evenly divided dataset balancing method can make the number of each group of data after balance processing more balanced,but it can’t guarantee that the characteristics of each group of data are sufficiently similar.（3）Construction of the aggressive driving behavior recognition model based on ensemble learning.Firstly,base classifiers are built on the training set obtained by dataset balance processing based on long short-term memory,gated recurrent unit,and temporal convolutional network,respectively.Then,based on the Max Rule,the Product Rule,the Majority Vote Rule,and the Sum Rule,the base classifiers built based on the above three deep learning methods are combined into ensemble classifiers for recognizing aggressive driving behavior.（4）Validation of the aggressive driving behavior recognition model.Firstly,classifiers are directly built on the imbalanced dataset based on 3 deep learning methods.Then,the testing set is used to validate the aggressive driving recognition performance of each ensemble classifier and the classifier.The results show that the ensemble classifiers built on the datasets obtained by the evenly divided dataset balancing method have better recognition performance of aggressive driving behavior,the LSTM ensemble classifier based on the Majority Vote Rule has the highest accuracy rate of90.29%,and F₁-score of 90.39%,which has the best performance for recognizing aggressive driving behavior.This research provides a new idea for research on driver behavior modeling.The theoretical reference for the improvement of advanced driver assistance systems and even the realization of personalized driver assistance systems,anthropomorphic automatic driving,and intelligent connected vehicles is also provided.