An anti-occlusion tracking algorithm

When¹ the calculated maximum filter response score is low, traditional correlation filter-based trackers can't determine whether it's drastic target appearance change or occlusion. They will take it as drastic target appearance change and introduce plenty of wrong training samples when target being occluded. In traditional correlation filter-based tracking algorithms, the original filter and the newly trained filter are fused together to approximate their weighted combination. So, these wrong training samples will pollute the original filter and this process is irreversible. In the proposed algorithm, when trackers can't determine whether it's drastic target appearance change or occlusion, target's bounding boxes' feature maps in that period will be assigned to a candidate set as candidate training samples. An independent filter is trained by training samples in the candidate set. This filter and the original filter work together to locate the target in subsequent frames but they are not fused. Their maximum filter response scores are recorded and analysed to determine whether the target has been occluded or not. Finally, these candidate training samples are fused into the original training samples set or be abandoned according to the confirmed situation.