Study On Dangerous Driving Behavior Recognitoin Based On Deep Learning

Dangerous driving behavior is an important cause of traffic accidents.By identifying the driver's dangerous driving behavior and warning,it can effectively reduce the risk of traffic accidents,which has important social significance and application value.In recent years,with the breakthrough progress of deep learning technology in the field of computer vision,the application of deep learning and computer vision technology in the identification of dangerous driving behavior has achieved good results,and has become a research hotspot for scholars at home and abroad.However,on a low-cost embedded device based on a monocular camera and limited computing power,how to effectively and effectively recognize the driver's dangerous driving behavior in real time still needs to be explored in depth.This paper conducts research based on deep learning methods.First,a human pose estimation method based on a convolutional neural network model Mobile Net V3 is proposed to obtain the joint point coordinates of the driver's upper limb skeleton.Then,the action recognition method of recurrent neural network based on skeletal sequence is researched,and a spatio-temporal simple recurrent unit(GCA-ST-SRU)model of global context attention is constructed to identify dangerous driving behaviors.Finally,the proposed model and method are programmed on an embedded platform to develop an embedded lightweight dangerous driving behavior recognition system.The main work and research results of this article are as follows:First of all,in order to solve the problem that the current deep CNN model takes a long time to perform pose estimation on embedded devices due to the deep layers,many parameters,and large calculations,it is beneficial to the speed of the Mobile Net V3 model.Mobile Net V3 pose estimation method.Testing on the LSP dataset and the upper limb pose dataset,the experimental results show that compared with the existing deep CNN model,the proposed method achieves higher accuracy while greatly reducing the number of parameters and improving the operation of the algorithm.speed.Secondly,a novel GCA-ST-SRU model is constructed to solve the problems of reliance on the calculation order of the recurrent networks such as LSTM,ST-LSTM,and GCA-LSTM,which leads to slow inference speed and low calculation efficiency.This method inherits the advantages of SRU training and fast inference speed.It can simultaneously model the temporal and spatial dependencies of joint points,and introduces an attention mechanism to selectively focus on joint points with high information content.The evaluation was performed on the public UT-Kinect action data set and SBU-Kinect interactive data set.Experimental results show that the GCA-ST-SRU method achieves a good balance between accuracy and real-time performance.Finally,the Kinect depth camera is expensive,has high cost and is difficult to popularize in practical applications,and the current research on dangerous driving behavior recognition based on monocular cameras is mainly focused on monitoring the driver's head,ignoring the driver's limbs.In this paper,an embedded lightweight dangerous driving behavior recognition system based on a monocular camera is developed in this paper,realizing real-time recognition of 8 behaviors.Experimental results verify the effectiveness of the proposed algorithm in practical applications.