Moving Object Detection And Tracking Based On Independent Component Analysis

Intelligent Transportation System (ITS) is one of the most important domains of computer vision. As more and more cars running in the city and highway, the increasingly worse traffic and the urgently demand of intelligent transportation control system, the ITS becomes one of the most popular research domains in and abroad China, and has a wide range of applications in the future. After thirty years' research, great achievements have been made in China. But it is still not completed and needed to be perfected, especially in some key parts such as moving object detection and tracking under no occlusion or occlusion cases.This thesis focuses on two crucial parts of ITS: moving object detection and tracking. There are some drawbacks in the traditional methods of moving object detection and tracking, such as heavy computational load, easily disturbed by noise and trembling and so on. In this thesis, independent component analysis (ICA) and independent component analysis with reference (ICA-R) algorithms are introduced to moving object detection and tracking creatively, a novel and stable moving object detection and tracking method was proposed, and an application of ITS was designed and developed based on this moving object detection and tracking method. The concrete work of this thesis can be divided into three parts:In the moving object detection part of ITS, video frames are regarded as mixtures of the independent sources such as background image and the foreground image which consists of the moving objects. With ICA algorithm, the independent sources could be estimated, and then the moving object could be detected and extracted from the video frames. Because of the statistical characteristics of ICA, the moving object detection method do not need a highly real-time background and the moving object can be detected accurately even in a severely changed illuminating environment.The video frames are taken as linear mixtures which consist of the background image and foreground image which contains the moving objects. Moreover, moving object tracking could be understood as separating an interested source from the mixtures while the priori information was already known. It seems that the problem of moving object tracking could be solved by ICA. The ICA-R algorithm is introduced to solve the problem of moving object tracking. Taking the invariable moment of the detected object image as reference, the right moving object which shares the same characteristics with reference could be separated from the video frames and be pitched in each frame by ICA-R algorithm. To address the difficult problem of non-totally occlusion, the detected object is first divided into two or four parts averagely, and then the unoccluded sub-parts can be tracked through ICA-R algorithm in which the invariant moments of the sub-parts are used as the references. Then the tracking of the whole moving object could be realized.The experimental results demonstrated the effectiveness of the proposed method.