Research On Driving Behavior Recognition Based On Smart Car Sensors

With the development of China’s economy,a large number of people are pouring into ke y cities and the number of cars is continuous increasing,which lead to frequent traffic accide nts and traffic violations every year.It is an important topic for intelligent transportation syst ems to reduce the rate of traffic accidents.At the same time,it is very important to obtain the status information of the vehicle timely in the intelligent transportation system.The occurre nce of traffic accidents is often closely related to the driver’s driving behavior,so we will con duct a study on how to identify the driving behavior accurately.There are several problems in the previous sensor-based driving behavior study:(1)Onl y single sensor is used in the study,and the single sensor data has limited ability to identify m any driving behaviors,and the accuracy is not high.(2)In the past,the method of hand-desig ned features is often used in the process of identifying driving behaviors if there is multi-sour ce sensor data.Since the dimension of features is as high as dozens,the calculation is very cu mbersome,it is greatly restricted by the professional field,and the ability of the convolutiona l neural network to locally extract features is not used.(3)At present,research on multi-sourc e time series sensor data focuses on the extraction and modeling of time series features and ig nores the connection between feature channels and the modeling and analysis of long sequen ce data.Therefore,we use the multi-source sensors of smart cars to identify driving behavior s.Two neural network models for driving behavior classification are mainly proposed based on the end-to-end idea.We combine the advantages of FCN(Fully Convolutional Network)and GRU(Gated Recurrent Unit)neural network models,and propose a GRU-FCN driving behavior identification neural network based on channel attention.The neural network uses the FCN network to replace the traditional manual extraction of sensor time series data features,and uses its convolution kernel to automatically extract local features on the time series data.The feature channels after convolution has different importance,so a squeeze-excitation block is added to the model to enhance the feature channels with strong expressiveness and inhibit the feature channels with weak expressiveness;at the same time,the GRU model is applied to the original data in order to obtain a longer,global time series feature representation.After the two are fused,the classification results of the model are obtained through Softmax function.It has the fixed size of the convolution kernel and the limited receptive field in the FCN network,and the single-layer GRU model has limited ability to extract long time series features.Therefore,based on the improvement of GRU-FCN model using channel attention,the TCN-ANLSTM(Temporal Convolutional Network-Attention Nested Long Short-Term Memory)is proposed.The TCN model has the characteristics of causal hole convolution,which can freely control the size of the receptive field and use the residual structure to prevent network degradation.the NLSTM can memorize longer time history information than the ordinary GRU model.At the same time,the time step attention mechanism is used for each time series feature to improve the classification performance.Finally,the classification accuracy of the model on the dataset is 93.75%,the precision rate is 93.1%,the recall rate is92.62%,and the F1 score is 92.74%,all of which are better than the comparison models.This paper develops a driving event assistant detection platform,deploys the model in the edge computing device,can detect the behavior state of the vehicle during the movement process,and has high application value.