Research On Key Problems Of Moving Object Detection In Natural Scene

Vision is the most immediate and main source of information acquired by people from external world. Going along with the rapid development of science and technology, especially the popularization of computer application, the computer vision technology has become a hot research issue for the scholars. Moving object detection is the basis for many fields in the computer vision. Whether segmented the moving object effectively plays an important role for the results in the intelligent video monitoring, moving object tracking and gait recognition applications. The importance of moving object detection determines that the study of moving object detection in natural scene has important theoretic significance and practical values.In order to take an in-depth study on some of the key technologies for the moving object detection in natural scene, the thesis combines the motion detection, stereo vision, fuzzy mathematics and some other related theories in different fields. It aims to solve the key problems such as dynamic background, cast shadow, intermittent motion, camera jittering, three dimensional spatial location for the object and so on. The main contents and achievements of this thesis are as follows:(1) According to the problem of dynamic background, a moving object detection algorithm based on multi-scale Gaussian pyramid model is proposed. A high and low double thresholds background difference operation is used to overcome the contradiction between object over-segmentation and noise by single threshold. All the thresholds are obtained automatically. The fuzzy mathematic theory is also introduced. The color of the pixel, temporal, spatial and locality are fused to optimize the object detection results.(2) According to the problem of cast shadow, the method of multi-scale, high and low double thresholds are brought to improve the traditional GMM algorithm, and the IGMM is proposed to obtain motion mask and overcome the dynamic background noise. The computational color model is used to overcome the influence of illumination change and light shadows. For the darker parts of the shadow, the optical properties, the region location properties and the edge features of the moving mask are combined to achieve the purpose of eliminating motion shadows.(3) According to the problem of intermittent motion, an algorithm based on complementary model is proposed to detect moving object. Two groups of IGMM models own different update rates are used, one group of high update rate models are used to obtain the information for the removed object, and the other group of low update rate models are used to obtain the information for the static object. By comparative analysis for the mask properties, the intermittent motion detection has been achieved. It can recognize the object stopped for a long time and still object that suddenly start moving scenes effectively.(4) According to the problem of camera jittering, the feature point matching theory in stereo vision is introduced, a background adaptive method through feature points matching between sequences is proposed. The offset caused by camera jittering can be compensated from some pair of stable matched points, recover the background matched with the current image, and then overcome large error detection regions caused by camera jittering in natural scene. Because the jitter amplitude is not very much, compared with the traditional stereo match, the matching error is small in this application.(5) According to the problem of spatial location for the moving object, the thesis combined the stereo vision technology with moving object detection in a single video sequence, a stereo motion analyzing system is designed to achieve the goal of spatial location. It expends the application range for the moving object detection technology in a single video sequence. According to the problems in the course of transition from monocular to multi-view, a solution is proposed based on the hardware. Combined the stereo lense with the high-speed camera, the problem of chip consistency, camera synchronization and the motion ghost can be overcome. Because the stereo images come from one image, the traditional moving object detection method in a single video sequence can be used directly to obtain motion mask.The thesis carries on deep research on the key problems of moving object detection in natural scene. According to different problems, the corresponding solutions are given. The proposed algorithms are verified with the public test image and compared with several state-of-the-art algorithms qualitatively and quantitatively. Experimental results demonstrate that the proposed algorithms in the thesis can achieve moving object detection in natural environment effectively. The detected effect evaluation parameters have a certain advancement.