Correlation Tracking via Mask and Multi-peaks Re-prediction

At present, most of the classic Discriminative Correlation Filter (DCF) use rectangle to track targets. Although this method of selecting targets by rectangle is convenient, the shortcomings are also particularly obvious. For example, DCF may mistake some background information as target. Moreover, the DCF simply uses the linear method to update the target template, which will introduce interference into the model. In order to solve the above problems, we propose a mask and multi-peaks re-prediction correlation filter algorithm. First, the segmentation algorithm is used to extract the edge information in the target region, and the target mask is made according to the edge information. In this way more accurate target information is obtained. Secondly, when the target is interfered, the filtered response graph will fluctuate drastically. We compare the similarity between the target of the first frame and the alternative region that made through multiple peaks to select the most accurate and effective target, and use this target to update the model. We evaluated the proposed algorithm on the OTB-2013 dataset and UAV123 dataset, experiments show that our algorithm can better deal with background interference and has better performance than other advanced algorithms.