Object Tracking Algorithm Based On Image Feature

Object tracking is an important branch of computer vision. It has been widely used in the fields of video surveillance, intelligent transportation system, human computer interaction, military reconnaissance and robot vision navigation. The purpose is to achieve the target tracking and positioning by a certain similarity measure and matching search algorithm. Many target tracking algorithms have been proposed during many years of research. However, there are still some problems in the target tracking technology, which includes scale changes, severe occlusion, rotation and fast motion. Therefore, Shift Mean and SIFT algorithm are combined together in this thesis. In order to reduce the influence of background or noise to feature extraction, a new method of target tracking based on feature matching is proposed, which is based on Shift SIFT and Mean feature matching.At first, the thesis studies the classical Mean Shift algorithm of object tracking, and then the popular Kalman filtering and its application in object tracking is also studied. Because the Shift Mean algorithm is based on the color histogram of the target region and the candidate target region, and the kernel function window size of the Shift Mean algorithm is kept constant in the tracking process, so the Shift Mean algorithm can not effectively track the target when the target appears scale changes or moves quickly and suddenly in the process of tracking. In order to overcome these shortages, the target tracking method based on the combination of them is also researched.Secondly, this paper also outlines a variety of commonly used point feature detection algorithm. Due to SIFT features have the invariance of scale, rotation, affine transformation and illumination changes. They are stable for noise and suitable for fast and accurate feature matching.Therefore, the target tracking algorithm based on SIFT features is also studied in this paper. In a certain extent, the algorithm can deal with the effects of scale transformation, illumination, partial occlusion and noise on moving target tracking. However, when the moving target has a serious occlusion or the feature points detected in the candidate target region is weak, the algorithm may appear to have a large tracking error and can not track the target. Finally, in order to obtain better tracking results in these complex cases, a target tracking method which combines the two methods effectively is proposed in this paper.Compared with the traditional target tracking algorithm, the proposed algorithm has the following advantages: firstly, the initial position of object can be obtained by SIFT feature matching, which can overcome the effect of rotation, zoom and occlusion on object tracking; secondly, the accurate position of object can be obtained by Mean Shift,which can improve the accuracy of object; thirdly, in order to guarantee the stability of feature extraction and feature matching, SIFT feature matching of the dynamic update feature library is adoptted in this method, which greatly improve the precision of matching; finally, the tracking mode is determined by calculating target occlusion. If the occlusion factor is greater than the threshold, it will go into the occlusion tracking mode, otherwise it will update the template and the feature library, and go into the tracking of next frame. The experimental results show that this algorithm has good tracking adaptability to the target occlusion, fast moving and scale changes.