Study On Moving Object Detection And Tracking In Video Sequence

Moving object detection and tracking employs the methods of computer vision and video processing to automatically analyze the image sequence from video input device. Its purpose is to detect, locate and track moving object and provide video understanding and scene comprehension with object and analysis basis. It spans many subjects including computer science, image engineering, pattern recognition and artificial intelligence, etc. With the rapid development of algorithm theory and hardware technique, object detection and tracking technology has already been applied in industrial fields,such as navigation weapons, scientific exploration, and civil fields, such as community monitoring, traffic flow, driver assistance, and human-computer interaction, etc. It will be used more widely.As to video object detection and tracking, there are still many problems no matter in theory research or in applications. Such as how to balance between the integrality of object and the accuracy of boundary during object detection, and how to adapt scale, template, speed and trajectory changes during object tracking. Large numbers of researchers have devoted themselves in the area. Based on the current research, our study is carried out for these issues. The main work can be summarized as follow:In the aspect of moving object detection, for the demand of accuracy and integrality, a spatio-temporal joint detection method based on markov random field (MRF) was proposed. In spatial detection, a watershed segmentation algorithm with Mean-Shift mark constraint was presented. It used the clustering property of Mean-Shift to find the probability density center in both image and feather space. Then the visual area of concern was marked based on the information from both space and was regarded as the constraint condition of watershed segmentation. This method avoided over-segmentation. Based on the simplified S-TMRF model, the posterior energy function was defined. After that, the information from both temporal space and spatial space were integrated in MRF-MAP framework to detect moving object. The result is more robust and precise.In the aspect of moving object tracking, to improve deficiency that the kernel bandwidth of Mean-Shift is not changeable which makes it difficult to tracking object with changeable scale, a novel adaptive scale updating algorithm based on boundary force was presented. Based on the analysis of the target model, the boundary force was introduced, which added feature matching constraint of pixels near object boundary to Mean-Shift tracking. The adaptive algorithm could locate the target's position and could adjust the bandwidth of kernel-function according to the estimated scale. Compared with traditional three-step method, our algorithm reduced the computation and computing complexity and could track the object with large-scale changes more stably.How to determine the tracking status was studied here. Based on the analysis of the relationship between object features and background features, the feature enhancement function was introduced and the novel background template was constructed. During tracking process, with the comprehensive analysis of similarity coefficients of candidate object and the templates, the proposed algorithm could accurately judge the tracking status and the cause of interference, then take corresponding template updating strategy. This method could make the judgment of tracking status more precise because of the analysis of the interaction between object and background.The particle filter algorithm was studied to track fast moving object with complex trajectory. Aiming at the problem that particle filter requires many particles to approximately describe state of object, which is more time-consuming, the Mean-Shift algorithm was used to converge the particles to area of real state before re-sample. It made distribution of particles more reasonable. Because the particle description became more rational, the number of particle required was reduced and the tracking efficiency was improved.According to the indoor environment, an object surveillant test platform was designed and constructed based on our algorithm. The test platform used TI's high-performance DSP TMS320DM642 as the core chip. Optimization has been carried out on three levels such as algorithm structure, code structure and chip function according to the property and feature of hardware system. In the system design, the host-guest structure was used to share the computational load. This platform made full use of the computational power and programmability of DSP。It was reconfigurable and was suitable for many kind of image processing.