Research On Object Detection And Action Recognition In Railway Visual Surveillance

With the development of China's railway informatization,surveillance cameras have become increasingly popular and the scope of monitoring has expanded.The intensity of work of monitoring personnel continues to increase.Traditional methods of relying on artificial viewing of surveillance video are not only time-consuming and labor-consuming,but also can easily lead to false detections caused by visual fatigue.This requires that machine vision technology can play a role in railway video surveillance,and monitor videos with high efficiency and accuracy for analysis.The purpose is to use machine vision technology instead of human eyes to study detection task and action recognition in railway video surveillance.The main research work of the dissertation is as follows:(1)Introduce the detection technology in railway video surveillance,analyze the structure of railway video surveillance system,and give the necessity of applying intelligent surveillance technology to railway monitoring.(2)Starting from the traditional frame difference method,optical flow method and background statistical model,we studied the moving object detection algorithm.Aiming at the topic that Gaussian Mixed Model can't extract a clear object,a detection method combining gradient image with Gaussian Mixed Model is designed.The solution to the problem of voids and shading that are easy to appear in foreground extraction is proposed.The features based on HOG and LBP are designed.Combined pedestrian detection model.(3)Analyze the network structure,loss function and network optimization algorithm in the aspects of artificial neural network.In order to solve the problem of low robustness of traditional detection methods,a railway object detection model based on Faster-RCNN was designed by analyzing artificial neural networks based on regional candidates.Finally,a railway target detection model was trained and used on self-made data sets.After training and verification,experiments show that the network can accurately detect small objects,and has strong robustness to the light environment.The average accuracy rate is 81.1%.(4)Analyze the abnormal action under different scenarios in railway surveillance video,design a prototype of the identification system for abnormal action in railway surveillance video,and propose a method based on 3D convolution in the process of action recognition for artificial design features.Neural network action recognition method,and modify the structure,increase the depth of the network,and modify the length and size of the input image,and finally design experiments to compare the impact of different sampling methods.Experiments show deeper network structure and input length can improve the accuracy of the action recognition.The average accuracy rate on the KTH is over 93%,which is about 3% higher than Ji's network.