Research On Distracted Driving Behavior Recognition Based On Deep Learning

With the rapid growth of car ownership and traffic flow,the total number of road traffic accidents that occur every year in countries around the world remains high.A large number of traffic accidents caused thousands of casualties,brought a painful blow to related families,and caused huge economic losses to society.Plenty of traffic accidents were caused by distracted driving.In recent years,research on the recognition of distracted driving behavior based on convolutional neural networks has made certain progress,but the existing algorithms have large weight files and long calculation time,and the recognition accuracy still needs to be improved.It is necessary to optimize and lighten the distracted driving behavior recognition algorithm.In order to solve the problems of the existing distracted driving behavior recognition datasets being small in public number and low marking quality,this thesis collected data on its own to construct a distracted driving behavior recognition dataset.This thesis considered gender and lighting conditions when collecting data,and collected distracted driving data of20 drivers whose male to female ratio is 11:9 under different lighting conditions.In order to ensure the accuracy of the mark,the mark was checked after labeling is finisded.In the division of training set and test set,based on the driver rather than scrambling all data together to avoid random division that may lead to overfitting of model training.The constructed dataset contains 28690 images of ten types of actions.In view of the small differences in distracted driving action and the low accuracy of ordinary convolutional neural networks for classification of subtle differences,the RFB-HBP model was proposed based on the receptive field module RFB and the fine-grained classification method hierarchical bilinear pooling model HBP.The model applied the finegrained image classification method to the recognition of distracted driving behavior.Considering the problem of the small receptive field of the convolution kernel used by the HBP model feature extraction backbone,the improved receptive field module RFB which was modified by dilated convolution with large eccentricity was applied to the HBP model feature extraction backbone.Compared with the HBP model,the accuracy of the proposed RFB-HBP model is increased by 3.03% to 93.26%.The Grad-CAM visualization heat map shows that the RFB-HBP model does pay attention to the key areas of distracted driving behavior.Distracted driving behavior recognition has strict requirements for real-time performance.In view of the poor real-time performance of current distracted driving behavior recognition algorithms and excessive model size,this thesis proposed a lightweight method for distracted driving behavior recognition.In view of the large amount of calculation of 1*1 convolution in Mobile Net V3 model and the problem of partial bottleneck structure without residual connection,ghost module was used to replace 1*1 convolution layers,and bottleneck structure with branch which is similar to residual mapping was used to replace bottleneck structure without residual branch,this thesis proposed a lightweight recognition model SMobile Net.Applying the knowledge distillation method,RFB-HBP was used as a teacher model,and SMobile Net was used as a student model.The former guides the training of the latter.The experimental results show that the two versions of the SMobile Net model are more accurate than the corresponding Mobile Net V3 model.At the same time,the model calculation and weight file size are smaller.Knowledge distillation improved the inference performance of the SMobile Net model.The performance of the embedded verification application shows that the two versions of the SMobile Net model have achieved the effect of real-time detection.