Vehicle Driving Behavior Recognition Based On Deep Learning

Vehicle Driving Behavior Recognition(VDBR)has been a challenging issue in the field of the vehicle driving and traffic safety.Considering the complex and diverse sections(highways,urban roads,country roads,etc.),and the differences in driver's driving habits,it is difficult to accurately and timely analyze and identify driving behavior information.One of the main methods of vehicle driving behavior recognition is to classify based on inertial sensor data.Most of the research uses smartphones with built-in sensors as data sources.Although using a smartphone as a data source is more flexible,there is more data interference.Correspondingly,other studies have used dedicated independent sensors.Most researchers use statistical machine learning methods,such as Support Vector Machine(SVM),k-Nearest Neighbor(KNN),Hidden Markov Model(HMM)and other algorithms for classification and identification.Few studies use deep learning solutions to solve this problem.Compared with the traditional machine learning algorithm,the deep neural network-based method can improve recognition accuracy greatly.However,the data collected by motion sensors have special correlation characteristics,and no scholars have proposed a special deep neural network to solve the problem of recognition accuracy.This paper presents a VDBR solution based on a six-axis inertial sensor.The method uses data collected by onboard sensors for classification by a deep learning algorithm.The main contributions are introduced as follows:(1)In order to solve the problem that the model is prone to overfitting due to the small amount of data,this paper proposes a Jointed Data Augmentation(JDA)algorithm based on inertial sensor time series data,including three sub-algorithms:Multi-Axis Weighted Fusion(MAWF)algorithm,Background Noise Fusion(BNF)algorithm and Random Cropping(RC)algorithm to build a data set that is more in line with the complex actual driving environment.(2)By analyzing the characteristics of vehicle driving data,a feature construction algorithm that can be applied to deep learning is proposed.This feature construction method improves the dimension of the original input features,provides more initial features,and improves the feature expression ability of the data set.The degree of dependence on the nonlinearity of the model is reduced,and the accuracy of the model is improved.(3)Propose a new neural network architecture,Multi-View Convolutional Neural Networks(MV-CNN),to extract features from four directions of input tensor,fully extract the spatiotemporal characteristics of the tensor.The direction of information transmission of traditional CNN models is improved.This model can be used for training,learning and classification of driving behavior.The evaluation results show that the data enhancement scheme increases the number of samples in the data set and makes the sample more balanced.The feature construction scheme weakens the dependence of the classification effect on the nonlinearity of the model and improves the accuracy.The MV-CNN network structure fully extracts the spatio-temporal characteristics of the data due to the introduction of multi-view convolution scheme.It can be seen from the experimental results.MV-CNN has the ability to mine the underlying characteristics of data.The prediction accuracy of this scheme can reach 95.29%,which is higher than the classical model.At the same time,the program has better generalization ability,reduces training variance and deviation,and improves the stability of the model training process.