Research On Detection Method Of Dangerous Driving Behavior Based On Video

With the popularization of private cars,traffic accidents caused by dangerous driving behaviors are increasing.Video-based dangerous driving behavior detection technology has important application research value.Dangerous driving behavior is spontaneous and accidental.Using traditional video behavior detection methods to detect driver behavior will inevitably cause information lag.Moreover,driver behavior contains timing characteristics,and efficient modeling of driver behavior is also an important challenge in behavior detection.It is not easy to implement video-based dangerous driving behavior detection technology.From the perspective of computer vision,this paper uses video frames,video spatiotemporal features,and spatiotemporal attention mechanisms as the starting point to carry out research on dangerous driving behavior detection.The specific research work of the thesis is divided into three parts:(1)Propose a dangerous driving behavior detection method based on video frames.First,determine the type of the current frame based on the low-level features of the video image and the historical decision information of the key frame;secondly,according to the type of the video frame,the high-level features of the video image are extracted in a corresponding manner,and the deep convolutional neural network is used to extract the high-level features of the key frame.The propagation method extracts high-level features of non-key frames;finally,based on state transition and LSTM network,two frame-based behavior detection methods are designed to detect the categories of dangerous driving behaviors in driving behavior videos.Experimental results show that the dangerous driving behavior detection method based on video frames can effectively deal with the detection delay caused by extracting the high-level features of each frame of video image.Under the premise of ensuring detection accuracy,the detection rate is increased by about 5FPS/sec.(2)A dangerous driving behavior detection method based on video spatio-temporal characteristics is proposed.First,the driving behavior video is disassembled in spatial and temporal streams,and the dual-stream data of driving behavior video is collected;secondly,based on the video space and time sequence,the spatial and temporal characteristics of the driving behavior video are extracted separately through convolution,and performed Feature fusion;finally,with Conv LSTM network as the basic unit,a network model based on Conv LSTM cascade is designed and used in actual dangerous driving behavior tasks.Experimental results show that this method can fully utilize the complementarity of video spatial and temporal features to extract the semantic information of driver behavior;secondly,the cascade structure can learn driver behavior features from shallow to deep,and improve the average detection accuracy of dangerous driving behaviors.By about 1%.(3)Propose a dangerous driving behavior detection method based on spatiotemporal attention mechanism.Obtain the salient features of driving behavior videos by introducing an attention mechanism.First,by detecting the driver’s contour,restrict the attention range of the spatial attention;secondly,use the LSTM network to calculate the spatial attention weight of each sub-region of the video image,and extract the salient spatial features of the video;at the same time,calculate the driving behavior based on the LSTM network The weight of each frame of the video,and extract the salient temporal characteristics of the video;then,based on the attention mechanism to guide the fusion of spatial and temporal features to calculate the salient spatio-temporal features;finally,the use of Conv LSTM cascade-based network to achieve dangerous driving behavior Detection.Experimental results show that by introducing a spatiotemporal attention mechanism,the network’s ability to describe various dangerous driving behaviors can be improved,and the accuracy of dangerous driving behavior detection can be improved by about 1.5%.