Remote Intelligent Surveillance System

The use of video surveillance system from simulative age, digital age to now network and intelligent age becomes more and more important in every walk of our life. And the intelligent application is deemed to the direction of its development. Intelligent surveillance aims at attempting to detect, identify and track human bodies from image sequences, and to understand and describe their behaviors.In this thesis, we focus on the problems of digital video capture, compression, transmission, storage, playing and intelligent image analysis. Our work concentrates on the tracking, segmentation and analysis of human motion in dynamic scene. The research includes following aspects:(1) Design and implement a Remote Surveillance System. The video captured from remote camera is compressed to MPEG-4 format, and transmitted to multiple computers in real time using RTP and multicast techniques. The implementation of the system is composed of many techniques, including video capture, compression, transmission and multimedia programming.(2) Introduce a motion detection algorithm. Nearly all the proposed algorithms start with the detection of the interested entities from the background scene. The detection is achieved based on self-adaptive background subtraction and shadow detection method, and improves the robustness of the system.(3) A hybrid tracking algorithm between region-based and feature-based methods is presented. Moving objects are detected and stated as connected binary regions, and each connected region is identified by its features. The goals of tracking are to determine when an object enters the filed of view, compute the correspondence matching between objects in continuous frame. The results show the robustness of the proposed algorithm to cope with multiple moving objects in dynamic scene.(4) A motion analysis algorithm is presented. The method is divided into two phases: motion detection and motion analysis. First, human motion mask is detected. Then, three extreme points are extracted which usually appear in the human boundary and having periodical behavior. The relative orientations of the torso are employed to define a metric for continuous human motion classification. The method shows excellent results in classifying human motion in terms of walking and running.(5) Present a moving objects segmentation algorithm. The segmentation is performed based on two stages: moving edge map extraction and moving object extraction. Moving edge map is used as a clue to identify object boundary, and it is extracted by combining edge points that are detected from frames difference and background subtraction. In object extraction stage, horizontal and vertical candidates are smoothed and noisy pixels are suppressed via a post-processing step. The results in different scene show that the segmentation is achieved quickly and correctly.