Research On Driving Behavior Recognition Algorithm Based On Deep Learning

With the popularization of private cars,the number of traffic accidents is increasing day by day.About one-fifth of these accidents are caused by distracted driving behaviors such as phone calls,typing,eating,and communicating with people.During the driving process,the driver's driving behavior is directly related to the safety of the entire vehicle,so it is necessary to identify the driver's behavior status.Current driving behavior recognition methods have low recognition status and low classification recognition accuracy.In addition,due to the small memory of embedded devices in public safety research,driver detection is still a difficult problem.The task of driving behavior state recognition can be regarded as a multi-category classification problem.Considering the latest advances in computer vision in predicting driver behavior and the advantages of deep neural networks to extract data features faster and more efficiently,this article attempts to study the best deep learning network architecture to accurately detect driver status through computer vision to monitor this type of dangerous driving behavior.In this thesis,we have studied the use of deep learning methods to automatically identify driving states in a single image(such as normal driving,calling,texting,drinking water,and talking to passengers),for which we propose a fast downsampling network(MF-Net),it is an efficient and accurate network,which is improved from a deep separable convolutional network.The key idea is to apply a fast downsampling strategy to a deep separable convolutional network.Downsampling,this design has the advantages of significantly reducing the calculation cost,increasing the information capacity and achieving performance improvements,and the small required storage capacity is conducive to future embedded system migration,etc.This provides a certain theoretical basis and Technical guidance route.The thesis mainly does the following work:(1)introduces the purpose and significance of driver behavior state recognition,the development status at home and abroad,and the future application prospects of the system.(2)The overall architecture of the driver status recognition system and the algorithms and technologies required for the architecture are explained.(3)Data processing and feature visualization.(4)Several common convolutional neural networks are studied,and a driver behavior classification algorithm based on deep separable convolution is constructed and designed to effectively detect the driver's talking on the phone,texting and other behaviors,and compared these algorithms.Experimental results show that our model compared with other CNN models such as VGG and ResNet50,MF-Net not only smaller but also shows better performance in driver status classification.