Visual tracking has made significant progresses both in algorithm design and practical tracking system development in recent years. It is one of the most difficult problems in computer vision and the research in visual tracking will bring great meaning to national defense and everyday applications. With the high processing speed in computer hardware, real-time tracking has the possibility to come into reality. There are many papers on such topics published every year. This dissertation analyzes the following problems in visual tracking area: initialization of tracking; algorithms comparison and detecting and tracking moving objects from a static background scene using color images. Current algorithms seldom pay attention to the initialization problem in tracking. Most of them do this by hand, while tracking automatically is on the demand. This dissertation discusses an object localization method based on statistical model and shows its effect through experiments. What not accommodates to the active research in tracking algorithms is the lack of systematical evaluation of tracking results. The evaluation difficulties lies in the fact that all the tracking algorithms are working under given tasks and there are no unified theory of visual tracking and no standard testing image sequences. So it is hard to establish a general evaluation guideline for all the algorithms, but people need such reference to choose proper method for a certain environment. Some comparison work of tracking algorithms has been done between region-based and contour-based methods in this dissertation. It focuses on the template-based region tracking method and CONDENSATION algorithm. Analysis is made in algorithms' assumption, prior information, computation and running speed. Conclusion is that template- based region tracking works well in simple background with the target being roughly constant from frame to frame, while CONDENSATION algorithm is especially useful when dealing with cluttered background and unknown change in motion. Color information is the most intensively used visual feature in virtue of its strong correlation with the underlying image objects or scenes. Compared to other low-level visual information, color is more robust with respect to scaling, orientation, perspective and occlusion of images. A new algorithm of detecting and tracking moving objects from a static background scene using color images is proposed. It is based on the comparison of color features to decide moving areas in the image. The effects of color features are analyzed in this dissertation. Another innovation in this algorithm is the choice of reference background image, since it doesn't require background learning process.
修改评论