Research On Long Term Object Tracking Method Based On Random Ferns

Object tracking is a fundamental issue in computer vision research. Long term object tracking in any complex environment, therein, attracts the interest of many researchers. In order to solve the object tracking in complex situation such as occlusions, partial and full changing illuminations, appearance change or fast motion. A novel object tracking framework which combine standard object tracking with object detection and object pattern update is proposed in the thesis. On the ground of the novel framework, further research on object tracking based on random ferns classifier is conducted, and the methods are validated experimentally.At first, the thesis discusses the characteristics of random ferns and discusses the role of random ferns classifier in the issue of long term object tracking. Optical flow tracker based on Lucus-Kanade method was selected as standard tracker. Then the forward-backward error filter was introduced to improve the results of standard tracker. The tracked points filtered by median filter remain to evaluate the object position. The experiment shows that the standard tracker failed to work in these situations such as occlusion, fast motion and appearance change. To re-initialize the object box, the object detection base on random ferns is combined with the standard tracker. The experiment shows this solution conquer the problem under the situation of occlusion and fast motion. However, the problem of object appearance change could not be voided. Moreover, the fail probability of tracker increased, since the object position would be re-initialized with the help of object detector.Ultimately, the whole structure and the process of the long term object tracking based on random ferns classifier system are proposed. The object pattern update method is proposed to solve the problem of tracker failed with appearance change. The evaluation index is also discussed before the final experiment. The experiment results show that the long term object tracking based on random ferns classifier method performs better than the resent methods in long term object tracking, and, moreover, the long term object tracking approach proposed in the thesis has a good overall performance.