Analysis Of Train Driver Behavior Based On Multi-scale Feature Extraction And Fusion

With the continuous development of railways,trains have become more and more important in economic and social life.Train safety has also become a top priority of railway work.Drivers play an important role in current train operations.Currently,supervision of safety driving is still dependent on manual inspection afterwards;efficiency is very low.Based on this background,this thesis uses computer vision technology to identify and analyze driver behaviors,and the recognition results are used to monitor the driver’s behavior and realize intelligent driver’s behavior monitoring.The main research contents of this thesis are as follows:(1)Video image enhancement of driver monitoring video.In this this,an improved Retinex algorithm is used to enhance low-illumination images in driver surveillance video.This algorithm enhances the brightness of the image while retaining more image details.(2)A method for driver’s 3D pose estimation and tracking based on monocular monitoring.The Open Pose algorithm was used to extract the two-dimensional pose of the driver,and the two-dimensional pose was reconstructed three-dimensionally by combining the camera imaging model and the human body model.The 3D Kalman filter algorithm is used to track the 3D joint points’ sequences to remove the noise caused by the pose estimation and 3D reconstruction.The experiments prove that method used in this thesis can effectively track and denoise the joint point trajectory.(3)A multiscale Profile-HMM action recognition method with dual matching states.The motion sequence is segmented based on the joint point motion trajectory,and multiscale motion features are extracted for the division points and motion fragments respectively.The improved K-means algorithm is used to cluster the motion features.And the uppercase and lowercase letters are used to encode the segmentation points and action fragments,respectively.The action sequence is converted into a code word sequence with alternate uppercase and lowercase letters.The encoded sequence is modeled and identified using Profile-HMM with double matching state.Through experiments on the MSRACtion 3D dataset and self-built dataset,the method in this thesis has significantly improved the recognition accuracy compared with the traditional HMM algorithm.(4)The driver behavior semantic space is designed to describe driver behavior semantically.The driver’s behavioral semantic space is constructed by combining the scene of the driver’s cab and the understanding of semantics,combining scene information,objects,and action information,and using a four-tuple approach to output driver’s semantics.