Robust visual tracking through hedging multiple features

Visual tracking is a challenging task because target objects often undergo significant appearance changes due to factors such as heavy occlusion, background clutter, deformation, and sudden movements. In this study, we propose a novel tracker that enhances discrimination between the target and the background by hedging multiple features, including handcrafted features and deep features extracted from various convolutional layers. A weak tracker is generated by applying correlation filters to each feature, and all weak trackers are hedged to create a stronger, more robust tracker. To achieve robust visual tracking, we employ a hedging approach that adaptively determines the weights of weak trackers by considering both the discrepancies between their historical and instantaneous performance and the variations among all weak trackers over time. Additionally, an adaptive window is utilized to emphasize the target region and mitigate boundary effects. Extensive experiments conducted on various benchmark datasets demonstrate the effectiveness and feasibility of the proposed approach. The source code for this study is openly available and can be accessed at the following link: https://github.com/xianyou916/hmft .