Design Of Driving Behavior Detection System Based On Deep Learning And Computer Vision

China’s automobile ownership has jumped to the world’s first place,and the number of drivers has grown rapidly.However,according to the World Health Organization’s 2020 report on road traffic accidents,about 1.3 million people die in traffic accidents worldwide each year,and the economic losses caused by road traffic accidents account for about 3% of their gross domestic product.Distracted driving and fatigued driving have caused more than 35% of all accidents.Therefore,it is necessary to identify and detect distracted driving and fatigued driving.This paper establishes a multi-view distracted driving behavior recognition model and a multi-task distracted driving behavior and fatigue driving feature joint recognition model based on multi-view learning and multi-task learning.Based on the fatigue characteristics of fatigue driving,this paper confirms the fatigue driving judgment strategy to realize the driving behavior detection system and contribute to improving China’s road traffic safety.Firstly,this paper establishes a multi-view driving behavior image dataset NMDA dataset and proposes a multi-view distracted driving behavior recognition model MMob Net.Based on the different requirements of the two tasks of distracted driving behavior recognition and fatigue driving behavior recognition,a multi-view driving behavior image dataset NMDA dataset was established from different perspectives,and the effectiveness of the dataset was experimentally confirmed.Based on the multi-view model MVCNN as the model basic framework and the proposed view attention mechanism VAM module,a multi-view distracted driving behavior recognition model MMob Net was established.After experiments,it was confirmed that the MMob Net model has superior performance to single-task models,with an accuracy rate of more than 10% higher.The MMobe Net model using the VAM module has an accuracy rate 1.02% higher than that of other similar multi-view models.The combination of views 1,2,and 3 in MMob Net’s view combination comparison can greatly reduce the data demand scale while slightly reducing the model accuracy.Secondly,this paper confirms the fatigue driving behavior recognition scheme and proposes a multi-task distracted driving behavior and fatigue feature joint recognition multi-task model MTDFNet.In the fatigue driving behavior recognition scheme,Retina Face is used as the face detector to collect face area images divided into upper and lower eye and mouth images to establish a fatigue feature dataset EMFD and establish a fatigue feature recognition model MEMNet.Based on the distracted driving behavior recognition model MMob Net and the fatigue feature recognition model MEMNet as the basic structure,a multi-task model MTDFNet is established using multi-task learning.Through experiments,it was confirmed that the loss function and loss combination algorithm used to train the MTDFNet model obtained excellent model accuracy.The accuracy rates of the two sub-tasks of distracted driving behavior recognition and fatigue feature recognition were91.22% and 93.30%,respectively.Compared with the corresponding single-task models,although the accuracy rate of the more difficult distracted driving behavior recognition task increased by3.7%,the accuracy rate of the fatigue feature recognition task decreased by 0.47%,which proves that multi-task learning can improve the performance of some sub-tasks.Finally,this paper completed the design and implementation of the driving behavior detection system.Based on the fatigue feature recognition model MEMNet,the fatigue feature determination strategy for fatigue driving was confirmed and tested,and the upper limit of the number of frames with closed eyes and the upper limit of the time with open mouth were obtained as key parameters.Using Pyqt5 as the system design framework,a visual interface and a driving behavior detection system containing the interface were established and system testing was performed.The results showed that the system can run effectively.