Research Of Object Detection And Tracking Algorithms In Intelligent Video Surveillance System

As people pay more attention to the public safety,video surveillance is being ap-plied to more and more scenes,resulting in massive video data.Traditional manual processing has gradually become infeasible,we urgently need to use the computer to carry out the unattended intelligent video surveillance.Intelligent video surveillance involves technologies such as object detection and visual object tracking,and object detection can be furtherly classified into the moving object detection and the object de-tection according to the given data.However,due to the dynamic nature of the real world,such as noise and illumination changes,moving object detection algorithms are often ineffective,and most of the object detection algorithms still suffer from the prob-lem of high computational complexity,visual object tracking algorithms nowadays also has got limited accuracy and low efficiency.We need to find better solutions to these problems.Therefore,it is of great theoretical and practical value for the research of key technologies in intelligent video surveillance.The major works of this thesis are summarized as follows.1.We firstly introduce some classic moving object detection algorithms,including background subtraction,coterminous frames difference,optical flow and background modeling,and give an analysis of these methods' advantages and shortcomings.2.We focus on the background modeling algorithms in the moving object detec-tion.The Gaussian Mixture Model(GMM)is a very effective background modeling al-gorithms.While as GMM is based on single pixel value,there are always some "holes"in its detection results.After we analyse the GMM and the Deep Convolutional Neu-ral Network(DCNN),we propose a moving object detection algorithm based on deep encoder-decoder network,and experiment results show that the algorithm we propose not only solves the"hole" problem in the original GMM algorithm effectively,but also greatly improves the robustness of the algorithm.3.In this thesis we discuss the general process of the object detection algorithms,and analyse key points in the algorithm.Facing the problem that most existing effective object detection algorithms always has high computational complexity,we intorduce the Faster-RCNN model,which is both accurate and effective.4.We study existing visual object tracking algorithms,and give the general process of the tracking algorithms.After analyse the GOTURN(Generic Object Tracking Using Regression Networks)model,which is used to track arbitray objects,we propose a deep regression network to overcome its shortcomings.And experiment results show that our model can be used to solve the tracking problem,and it can be easily integrated into any existing network model.Experiments show that the algorithms we proposed can be used in intelligent video surveillance system,and these algorithms can improve the overall performance and efficiency of the system.