A Research Of Driver’s Behavior Recognition And Analysis With Convolutional Neural Network

With the gradual popularity of motor vehicles all over the world,people are accommodating traffic accidents while facilitating their lives.According to the statistics of the World Health Organization,more than one million people die in traffic accidents worldwide each year,and nearly one-fifth of the accidents are caused by the driver’s distracted behavior.Driver safety behavior detection is to detect and warn drivers of unsafe behaviors in the process of driving motor vehicles,such as using mobile phones,smoking,drinking water and other unsafe actions.In the field of image detection,traditional machine learning performs manual feature extraction and feature matching on the images to be detected,and secondly uses support vector machines for classification and detection.However,with the diversity and uncertainty of background,illumination and recognition objects during driving,traditional classification methods have failed to achieve good recognition accuracy in different scenarios.With the gradual popularity of deep learning methods in the field of image recognition,image detection and classification problems have been greatly improved compared to traditional methods.Starting from the initial image recognition model,the continuous stacking of convolutional neural network layers can train a deep learning model with high representation characteristics.However,the continuous increase in the number of layers of deep learning models will be accompanied by a series of problems.At the level of computing resources,the increase in the amount of model parameters requires larger storage resources and consumes more disk space.The training and detection phases of the model consume a lot of computing resources.Under the same training result,the larger the parameter of the model,the generalization effect is often inferior to the model with small parameters.The lightweight model designed in this thesis is based on the convolutional neural network structure,using separable convolution kernels and batch normalization technology,so that the block structure has a variable amount of parameters,and secondly,the use of small convolution kernels to achieve dimensionality reduction,achieving end-to-end detection.This thesis proves through experiments that the model with three consecutive blocks has the strongest characterization ability.While having high detection accuracy,it overcomes the shortcomings of slow detection speed of existing models.A test accuracy rate of 95.15% was achieved on the public dataset,and the model size was only 2.5M,and the detection speed reached 51 frames/sec.This result exceeded most models.The website provides free image detection and gradually collects driver-related images to provide more data support for subsequent scientific research progress.The model has a detection accuracy of 85.31% and a detection speed of 51 frames per second on the truck driver surveillance video compiled in the laboratory.Here,based on the results,the algorithm module of the driver’s safety behavior detection system is designed,and the system is assisted in landing use.