Fast Long-Term Tracking via Probabilistic Regression and Target Center Regression

Abstract

The difficulty of long-term tracking mainly lies in the tracking failure caused by out-of-view, occlusion, and intra-class distractors. To deal with these difficulties, existing methods divide the tracking process into two parts, local tracking and global re-detection, and focus on designing robust distractor awareness mechanisms and global re-detection strategies. However, most state-of-the-art long-term trackers ignore the efficient design of the model and therefore may either be inefficiency or time-consuming. In this work, a real-time Long-term Tracking framework based on a local-global switching strategy is proposed, which utilizes Probabilistic regression and Target center regression to accelerate the long-term tracking process (named LTPT). To achieve reliable interaction between the local tracker and the global re-detector, LTPT introduces a probabilistic perspective to interpret the output of the local tracker. Additionally, this probabilistic regression approach benefits from the complementary use of attention mechanisms and deep cross-correlation to emphasize multiple hypotheses in the search region. To achieve fast re-detection, LTPT uses a rough target center regression to obtain a search region that may contain the target, and then refines the target position through a local tracker. Experiments on several datasets demonstrate that the method achieves comparable performance to the state-of-the-art algorithms.