Object tracking-by-detection under cluttered environments based on a discriminative approach

In many visual tracking applications, an object is first detected and a tracker is then trained to track the object only based on the one-shot information. Such mechanism is called tracking-by-detection and has become increasingly popular. This paper describes a discriminative algorithm, which requires no offline training and is adaptive to variations of the appearance of the target, for tracking-by-detection. A discriminative tracking model does not build an exact representation of the target but tries to find decision boundaries between the object and the background. In this paper, the classifier employed to distinguish the object from the background is trained by a boosting learning algorithm. To suppress undesirable drifting effect, weight saturation is incorporated into the boosting learning algorithm. Drifting effect which is commonly seen in many adaptive trackers is inevitable since each time the tracker is updated, some error is introduced. Tracking-by-detection will make the tracker even more unstable because of the imperfect detection. Experimental results show that our approach can alleviate drifting effect while the adaptiveness is still retained. Also our approach is comparable to other visual tracking algorithms in hostile situations such as occlusion and cluttered backgrounds.