Object tracking via Online Multiple Instance Learning with reliable components

Abstract Visual object tracking is a challenging and essential research problem in the field of computer vision. Recent years, many Online Multiple Instance Learning (MIL) tracking methods have been proposed with promising experimental results. These methods train a discriminative classifier under the boosting framework. The weak classifiers are learned from parts of the object and all classifiers are updated while updating the appearance model. However, due to irregular shape of object or occlusions, some components are not on object and should not be learned. On the contrary, a discriminative weak classifier learned from these components will mislead the tracker to drift away. To overcome this problem, we propose a novel online MIL tracking approach by updating with reliable components (OMRC). It keeps both background and object templates while tracking. By comparing current tracking result with two templates, we can get whether the pixels belong to object. The components which have a higher rate of pixels belong to object than a predefined threshold are reliable components. Moreover, in order to represent images better, we use HOG features and Histogram features instead of the widely used Haar-like features and design a new online weak classifier learning method. Experiments are performed on two challenging datasets including OTB2015 and Temple Color. Experimental results demonstrate the robustness of our OMRC tracker and the effectiveness of each component in the OMRC tracker.