Research And Application Of Key Technologies For Driving Distraction Recognition Based On Deep Learning

According to a global study by the World Health Organization,road traffic accidents kill more than 1.35 million people yearly.According to investigations,the proportion of fatalities caused by distracted driving accounts for about 10%of all traffic accidents.With the proliferation of in-car entertainment apps and smartphones in cars,there are more and more possible causes of driver distraction.Therefore,there is an urgent need for a real-time distracted driving recognition and early warning system to help drivers focus on driving tasks and reduce traffic accidents caused by distracted driving.The current research on distracted driving mainly uses the public driving posture dataset,based on the improved deep learning technology,to identify distracted driving behavior.However,the following problems still exist in the current research on distracted driving:(1)The public datasets of distracted driving research are small in scale,and the data collection scenarios are single;(2)The division method of distracted driving datasets is unreasonable and cannot cover The actual complex application scenarios;(3)the generalization ability of the distracted driving algorithm model is poor;(4)the current distracted driving model has a large number of parameters and high complexity,which hinders the terminal deployment research application of distracted driving.This dissertation first focuses on the problems of the current distracted driving datasets,such as few publicly available datasets,small data scales,and poor data diversity.It constructs a public large-scale and diverse distracted driving dataset and designs four The dataset division method in line with the actual application scenario,research on the improvement of cross-domain generalization of distracted driving recognition,the progress of the accuracy of distracted driving recognition and the lightweight deployment of distracted driving recognition models;this research is helpful for real-time detection of distracted driving behavior.The main work of the thesis has the following aspects:1.Aiming at the problems of a small amount of data,single scene,and poor diversity in the current public distracted driving datasets,the largest and most diverse dataset 100-Driver,is constructed.The dataset collected 100 drivers,four angles,21 distracted driving behaviors,and one normal driving behavior.About 79 hours of multi-modal natural driving data were collected during the day and night.After data cleaning,there are more than 470,000 good pictures of 1920× 1080 size.Compared with the published datasets,the 100-Driver dataset has scale advantages in the number of samples and drivers.It has various advantages in distraction categories,behavior styles,camera angles,vehicles,and data modalities.The construction of the 100-Driver dataset is conducive to accelerating the research progress of gesture-based distracted driving recognition.2.Given the unreasonable division of distracted driving datasets and the poor generalization ability of distracted driving,four kinds of dataset division methods for actual scenarios are proposed.A domain adaptive method is introduced to solve distracted driving Generalization Problems Across Domains and Datasets.The 4 dataset splits include traditional dataset splits without domain changes,and three challenging dataset splits with domain changes,namely cross-angle,cross-modal,and cross-vehicle dataset splits.Using the latest deep neural network method,many benchmark experiments and in-depth analyses have been carried out on these four dataset divisions,and feasible suggestions for actual deployment are put forward.Aiming at the problem of poor driving recognition ability of cross-domain analysis,we proposed introducing an adaptive domain method to solve the generalization problem of distracted driving.Experimental tests have significantly improved recognition accuracy in distracted driving cross-domain and cross-data scenarios.Cross-domain dataset division and experimentation are the first public proposals in the field of distracted driving.The cross-domain dataset division and the method of improving the generalization ability of cross-domain distracted driving recognition are conducive to the model’s generalization to actual complex scenes.3.Aiming at the problem of background redundancy caused by wide-angle cameras and the influence of camera angles on distracted driving recognition,a method for improving the accuracy of distracted driving recognition based on data enhancement and data fusion is proposed.Traditional image data enhancement technology has limited improvement in accuracy;at the same time,the original distracted driving image contains a large number of redundant background areas.It is proposed to expand the dataset by extracting the driver’s operation area to obtain an effective method for data enhancement.Experiments on the distracted driving dataset 100-Driver and AUC show that our proposed data enhancement method can effectively improve the accuracy of distracted driving recognition.Aiming at the problem that different camera angles have different recognition accuracy for the same category,a multi-angle fusion distracted driving recognition model is designed.The accuracy of distracted driving recognition is improved by fusing the prediction results from multiple angles.Experimental results prove that the strategy based on dual-camera angle fusion can effectively improve the performance of distracted driving recognition.In practical application scenarios,it can be considered to jointly enable dual-camera perspectives to improve the accuracy of distracted driving recognition.4.Aiming at the problem of deploying the distracted driving recognition model on the terminal side,a lightweight model method is proposed to optimize the network convolution block and the number of channels of the fusion filter.Through the sensitivity analysis of the network,the convolution block of the optimization network is assisted in improving the recognition accuracy.At the same time,the number of fusion filter channels is compressed to achieve lightweight requirements.While considering the recognition accuracy,the model parameters are reduced to improve the model recognition speed.The lightweight experimental results on three distracted driving datasets 100-Driver,AUC and StateFarm,demonstrate the effectiveness of the lightweight method in this dissertation.The proposed method is deployed on an embedded terminal for testing.The speed of driving behavior detection on the Kirin990 system is increased by 1.5 times,which meets the real-time requirements of distracted driving recognition.