Improvement And Realization Of Moving Object Detection And Tracking In Video Surveillance System

In recent years due to serious public safety situation, as a core technique in the security monitoring field, intelligent video surveillance is widely applied. With utilizing the following technologies to intelligently process video data, such as computer vision, pattern recognition, machine learning, data mining and so on, the objective of intelligent video surveillance is to detect out interesting objects and behaviors in the monitoring scenes, and then it can be realized that the interesting objects and behaviors are exactly recognized and classified, efficiently stored and retrieved. An intelligent video surveillance system can be generally arranged from low level to high level of its contents as follow: image preprocessing, background modeling of scenes, moving object detection, moving object tracking, object recognition and classification, behavior understanding and analysis, etc. Thereinto, the most fundamental methods are moving object detection and moving object tracking. Although they have went through years of research and development, but it has been shown that technologies of moving object detection and moving object tracking are currently still immature in the practical experience. Especially, the following problems have bring great challenges to their researches, such as dynamic scene, appearance variations, scale variations, pose variations, occlusions, etc. This dissertation puts forward several methods to discuss the problems, such as moving object detection based on the markov chain and image segmentation, moving object tracking based on fusion within particle filter and saliency detection and so on. The major contents and breakthrough points are as follows:Firstly, moving object detection is studied from scene understanding, background modeling, foreground detection, background updating, and then the method of moving object detection based on the markov chain and image segmentation is proposed. In order to emphasize the disturbance of complex and dynamic background, this method utilizes the extrinsic nearsightedness characteristic of frog’s visual system, and proposes a mechanism of fuzzy scene understanding to restrain the disturbance. By analyzing the temporal-spatial correlation of neighboring pixels in images, a background model based on temporal-spatial graph is built. And then a coarse foreground detection is realized by combining the background model with the markov chain, as well as a fine foreground segmentation with the method of image segmentation is obtained. This hierarchical process effectively improves the accuracy of foreground detection. Base on the foreground detection, the factors that influence the accuracy of background updating are analyzed, and a sampling method based on random selection and dynamic updating is proposed, so as to effectively simulate some periodic or pseudo-periodic background motion simulation in the scene. The experiment on open datasets proves that the proposed method can obtain effective and robust moving object detection when the above factors are explicitly considered.Secondly, moving object tracking is studied from state estimation, saliency of visual attention, the fusion of them, and then the method of moving object tracking based on fusion within particle filter and saliency detection is proposed. In order to emphasize the tracker drift, occlusion and background disturbance, this method introduces the human visual attention mechanism in object tracking in the framework of object tracking based on particle filtering. It uses the visual saliency detection to guide object tracking and object appearance online updating, as well as estimates background around the foreground object. The introduction of the visual saliency detection makes the method deal with the occlusion problem, and the visual saliency detection provides a supervision to object appearance online updating so as to process the tracker drift. The experiment on open datasets proves the proposed method can obtain effective and robust moving object tracking when the visual saliency detection is introduced.