Study On Distracted Driving Behavior Recognition Based On New Pre-training Network

Driver’s distracted driving behavior is a big hidden danger of traffic safety.If driver’s distracted driving behavior can be identified and corrected in time,it will effectively reduce the frequency of traffic accidents.In recent years,the driving behavior recognition method based on convolutional neural network has become the focus of domestic and foreign researchers,and has achieved good results in theory and application.However,compared with the new neural network models such as Inception V3,DenseNet and Mobile Net V2,the classical neural network models such as Alex Net,VGGNet and Res Net,which are widely used in the research of distracted driving behavior recognition,have the problems of low recognition rate and large model volume.This thesis focuses on "The recognition of distracted driving behavior based on the new pre-training network",and proposes the recognition method of distracted driving behavior based on the new pre-training network and the improvement of network structure,the main contents are as follows:First of all,this thesis compares the classic pre-training network models such as Alex Net,VGGNet,Res Net with the new pre-training network model such as Inception V3,DenseNet,Mobile Net V2 to explore whether the new network model is better than the classic network model in recognition accuracy and model size in distracted driving behavior recognition task.The experimental results show that the recognition accuracy of new network models are generally higher than that of classic network models,and the model volume is generally smaller than that of classic network models.Among them,the recognition accuracy of DenseNet121 is 94.83%,and the model size is only 30.8mb.Then,the original DenseNet121 network structure is improved by asymmetric convolution,which imitates the design idea of Inception V3 network..The 3×3convolution layer of the four Dense Block layers is replaced by the 1×3 and 3×1convolution layers,and the recognition accuracy of the improved DenseNet121 network on the test set of State Farm driving behavior data set reaches 96.07%.Finally,based on the improved DenseNet121 network training model,the distracted driving behavior recognition system is developed to realize the recognition and early warning function of distracted driving behavior.The system test results show that the system successfully completes the human driving behavior recognition in the cab environment and the driving behavior recognition in the test set pictures of the State Farm data set,and achieves 96.07% recognition accuracy and the recognition rate of 30.80 frames per second on the State Farm dataset,which shows the effectiveness and practicability of the system.