An Efficient Occlusion Detection Method To Improve Object Trackers

Occlusion is one of the challenging problems in visual tracking. Most of the previous works alleviate this problem by randomly sampled weak features, or analyze it by methods closely related with specific trackers. In this paper, we propose an effective mechanism to detect the occlusion status by random forests, and embed this method into object trackers with a common strategy. We divide the target region into some regular parts, and extract the pairing features within and outside the parts to encode the structure information of the tracked target. The random forests are online trained to discriminate the occlusion status of the parts using occlusion-dependent samples. Several challenging video sequences are used to verify the model, and it proves that our model is capable of recognizing the occlusion status during tracking. The performance of two typical state-of-the-art object trackers is improved by embedding this occlusion detection method.