| Abstract [eng] |
Visual object tracking is one of the most important tasks of computer vision. Many various algorithms and technologies have been developed for this task. In this thesis, task of object tracking and difficulties of it are described. Three groups of object tracking algorithms are discussed, while two of the algorithms - Optical Flow Lucas - Kanade and CAMShift - are studied in detail. The combined method of these two algorithms, which is implemented in visual recognition system, is then presented. The method dynamically selects more suitable algorithm for each recognized object. CAMShift is being selected for single-color objects, while Lucas - Kanade best suits for tracking multi-colored objects. SURF algorithm is used for initial object detection. If Lucas - Kanade method is selected, SURF algorithm is used for calculating points to track. If critical number of tracking points is lost, correction of points from recognized object model is applied. Proposed method enables recognition system to track both single-color and multi-color objects without initial learning, makes tracking more robust and occlusions-proof. |
