Image Recognition Based On Convolutional Neural Network Research On The Safe Driving Behavior Of Metro Drivers

Based on the action recognition of computer vision,this dissertation analyzes the current driving behavior of subway drivers.In order to determine the time boundary,action type and abnormal situation of subway driver’s driving behavior,this paper improves the convolutional neural network,and compares the improved method with mainstream action recognition and sequential action detection methods.Subsequently,a test was carried out on a subway line to verify the effectiveness of the improved method.The main research content and results of this dissertation include:(1)Propose an algorithm for action recognition with short time sequence and low computing power.In order to meet the real-time requirements of the scene,an efficient action recognition method is needed.By adding a filtering gate mechanism,the information of multiple frames of pictures in time sequence is merged into one picture,so that 2D convolution can also obtain the characteristics of the action in time sequence,which reduces the computing power required to perceive the time sequence dimension.By stacking the sequential 2D convolution structure,the range of perception on the timing is expanded,which satisfies the discrimination of the driving behavior of subway drivers in the short timing range.(2)Propose an algorithm for predicting the boundary of online long video actions.Simplified the collection method of dual-stream network features,using 3D convolution to build a spatio-temporal feature extraction structure,greatly improving the efficiency of the feature extraction stage.By improving the processing method of the convolution kernel,its perception range has the ability to expand,and its expansion range increases exponentially,which can satisfy the judgment of the spatio-temporal information of the long video.Through the combination of spatio-temporal features and the long-term spatio-temporal perception network,it satisfies the judgment of the driver’s action boundary in the subway online long video.(3)Research on anomaly detection algorithm based on target tracking.The method of object detection and target tracking is introduced,which solves the problem that traditional target tracking requires online learning.By building a backbone network for feature extraction,the same set of features can be used for the initial positioning and subsequent tracking of the target.By combining the backbone network,target tracking and object detection,a self-recognition multi-target tracking algorithm is constructed,which satisfies the self-recognition and tracking of the limbs of subway drivers,and then determines whether the driver is in an abnormal state of losing driving ability.(4)Through the combination of online motion boundary detection and high-efficiency motion recognition,it is judged whether the driver is operating in accordance with the regulations during the entire subway driving process.For the detection of emergencies,a self-recognition multi-target tracking algorithm is used to determine whether the driver’s limbs are in an abnormal driving section.(5)Collect relevant video data on a subway line,and train three neural networks respectively.Through the data analysis of the trial operation stage,this method effectively identifies the driver’s driving behavior,time of occurrence,and whether it is abnormal.In order to further verify the effects of the three algorithms,they were compared with the current mainstream motion recognition,time-series motion detection,and target tracking algorithms.The results show that the method in this paper is not less accurate than the mainstream algorithms,and the computational cost is far It is smaller than the current mainstream method,can be effectively deployed on the edge computing end and maintain real-time processing capabilities.