Research On Distracted Driving Behavior Recognition Based On Deep Learning

The driving risk caused by distracted driving is one of the main causes of traffic accidents.Detecting and warning drivers’ dangerous behavior can effectively correct drivers’ unsafe behavior,which is of great significance to promote driving safety.In order to improve the detection accuracy of driving distraction behavior and realize the lightweight and real-time detection of the model,this paper studies the distracted driving behavior detection algorithm by using deep learning technology,creates the driving distraction dataset,analyzes the risk information contained in the image,classifies the distraction features and locates and recognizes the object through convolutional neural network.The main research contents are as follows:(1)Aiming at the problems of redundant parameters and huge model of traditional detection network,an end-to-end lightweight network for driving distraction classification is built.Considering the different distraction categories of driving distraction dataset,the attention module is embedded in the feature extraction process,and the basic structure of EfficientNet model:MBBlock is improved for fine-grained recognition.In the process of feature extraction,the depth separable convolution is used to reduce the model parameters.The network is deepened by using a large number of residual structures,and the residual network ENet＿AT based on lightweight convolution is constructed through the repeated stacking of multiple feature extraction modules。 Aiming at the experiment of driving distraction dataset,CutMix is used for data augmentation,and the label smoothing method is used to train the model to make the model more robust.Comparing the training results with the traditional convolution classification model,it verifies the accuracy of the model on the dataset,and realizes the lightweight of the model.(2)Aiming at the problem that the classification model only focuses on the global information of the image and loss of location information,the semantic analysis network of driving distraction behavior is constructed based on the object detection theory.The dataset is created by labeling 17 distraction targets/ action area,and the driving distraction area is classified and located.In order to meet the accuracy and speed requirements of distracted driving detection,the YOLOv5 detection model is improved.the attention module is embedded to optimize the network feature extraction process,and the C3 block is used to reduce the complexity of the model.In order to improve the detection performance of small objects,a lower-level feature map is added in the feature fusion process,and Bi-FPN is used to optimize the feature fusion process.The SPPF block is used to improve the calculation speed of the model,optimize the prediction box positioning loss function,and use the Mosaic data augmentation method to improve the learning ability of the model to the data.Finally,the experiment are designed to verify that the improved YOLO＿Dectection’s performance is better than the traditional model,and the driving distraction risk action judgment criterion is formulated.(3)Considering the practical application scenario of early warning for driving distraction risk,according to the judgment criteria of driving distraction risk action,the driver behavior is detected by the model.The model established in this paper is deployed on the server and android app through Python flask and Android studio respectively,which realizes online and offline detecion for the driver distraction,and verifies the feasibility of the deployment of this method.