Discriminative Active Contour Tracking Algorithm Based On Fusing Multiply Features

Target tracking plays an important role in visual analysis and understanding of target motion, it's the intermediate-level vision part. Target tracking is to detect, localize and track moving object in video sequences captured by cameras. Contour tracking is a way of target express based on target edge profile. Due to the contour can be flexible to handle the the changes of the target size, and provide detail shape information of the target, thus become an important research direction in the field of target tracking. But there are still many theoretical and technical problems to be solved. In this thesis, the main research target is the condition that there exists similar in appearance between target and the background. Fusion target appearance information and depth information to realize the target contour tracking. The main work is as follows:First, extracting the apparent feature of the target and the background to establish discriminative model between target and background, classify the test image to obtain confidence map, which will be used as the external energy of contour in the Level Set contour tracking framework to guide the evolution of the contour.Secondly,in the previous work, combined apparent features and depth information, and proposed a layered tracking framework based on Level Set which is used for the problem that background and target is similar in appearance. First of all, in the first layer, initial contour evolution according to apparent information; Then, combined with the depth image, we decisions whether it needs depth information to guide contour evolution. If it does not need, the results of the first layer is used as the final tracking results. If need, the result of the first layer in apparent features will continue evolution under the guide of apparent features and depth information. In the second layer contour evolution, we proposed a local weights based on depth image, which is used to fusion apparent features and depth information.Finally, for a series of experiments, the proposed method in this thesis based on the fusion of multi-feature discriminative model active contour tracking algorithm, is verified under 3 different scene, they are respectively appearance similar between background and target, depth similar between background and target, both appearance and depth are similar between background and target. The experimental results show that the proposed active contour tracking algorithm based on multiple features fusion has good accuracy and robustness under the situation of neither appearance nor depth is similar between background and target.