Driver Behavior Recognition Based On Deep Learning

Driver behavior recognition refers to identifying the driver's abnormal behavior during driving the vehicle.With the development of technology and economy,automobiles have become an important part of people's daily life.It is of great significance and extensive demand to identify driver behavior and alarm in real time.However,there are relatively few commercial products and academic research in this area,and most of them are limited to the identification of fatigue driving.Based on the vehicle single camera video surveillance system,this paper applies the deep neural network to the driver behavior recognition task.A driver behavior recognition and warning system based on deep learning was designed and implemented which can identificate drivers' unsafe or uncivilized driving behavior,including yawning,snoring,smoking and making phone calls in real time.This paper mainly carried out the following three aspects:1.A driver behavior recognition algorithm model based on deep learning is designed.The four behaviors of the driver are classified by a single image.The MTCNN face detection network is used to extract the local and global features of the face,and the driver behavior is identified by the deep residual neural network ResNet.The algorithm can achieve the correct rate of 97% and above for all four behaviors;2.For the low-power computing scenario,by simplifying the network structure,the operation efficiency of the model is greatly optimized,and the calculation amount of the classification network is reduced.In order to optimize the accuracy and generalization ability of model simplification,CoordConv is used to introduce position information into the model,and the knowledge distillation in the field of model compression is used to train the network,so that the model still has a high accuracy rate for the four behavior classifications,which can reach 94%;3.The image sequence processing method is used to deal with the driver behavior recognition problem,and the context information is introduced into the behavior recognition model to improve the accuracy and generalization ability of the lightweight model.Based on this idea,two approaches were tried.One is to extract the sequence features of the image sequence based on the RNN-LSTM.The other is to use multiple images directly as the input of the network model.Experiments show that both improved methods perform better than the one based on single image.The experimental results under real vehicle scene data prove that the driver behavior recognition algorithm model proposed in this paper has a good accuracy performance.With on-board real-time warning system,the driver's behavior algorithm has high practical value,and deserves further study.