Time-driven Temporal Block-based Correlation Filter for Visual Tracking

Traditional correlation filter algorithms, while pursuing high tracking performance, overlook the reversibility of tracking and fail to utilize the potential for performance improvement through inter-frame information. The existing methods have introduced a temporal block-based strategy, which allows the filter to inherit the discriminability of previous filters while adapting to more drastic object appearance changes and better utilizing interframe information. However, in order to address boundary effects in complex environments and abnormal response maps caused by model degradation during filter update processes, this paper proposes a time-driven temporal block-based correlation filter tracking algorithm called TB-TDCF. It combines the strategy of temporal blocks with a time regularization term and incorporates it into the base model to ensure that correlation filtering maintains similarity with previous filters during model updating. This procedure effectively suppresses degradation and updates anomalies caused by abnormal response maps, thereby improving accuracy and robustness of the tracking model over long duration tracking processes. Through extensive comparative experiments on datasets such as UAV123, DTB70, and UAVDT-M, this paper validates the superiority of our algorithm in terms of success rate and precision accuracy.