LAUSR

In recent years, correlation filter (CF)-based tracking methods have demonstrated competitive performance. However, conventional CF-based methods suffer from unwanted boundary effects because of the periodic assumption of the training and detection samples. A spatially regularized discriminative CF (SRDCF) has greatly alleviated boundary effects by proposing the spatial regularization weights, which penalize the CF coefficients during learning. However, the SRDCF utilizes a naive decaying exponential model to passively and fixedly update the CF from the previous results. Therefore, if the target meets with occlusion or is out of view, the SRDCF may encounter over-fitting to the recent polluted samples, which may lead to tracking drift and failure. In this paper, we present a novel CF-based tracking method to resolve this issue by dynamically and adaptively correcting the weights of learning CFs and fusing them together to promote a more robust tracking. Thus, if the recent samples are inaccurate in the case of occlusion or are out of view, our method will down-weight the corresponding CFs and vice versa. Moreover, in order to decrease computational complexity and ensure memory efficiency, we extract the key CFs from the previous frames to remove redundant CFs under the contiguous frame indexes constraint. Thus, we do not need to store all CFs and decrease computational burden. Benefiting from the extraction and enhancement of CF, our method improves the tracking precision on OTB-2015, VOT-2016 and UAV123 benchmarks and achieves a 56.0% relative gain in speed compared with the SRDCF. The extensive experimental results demonstrate that our method is competitive with the state-of-the-art algorithms.