FSTrack: Visual tracking with feature fusion and adaptive selection

Abstract

Visual object tracking represents a critical research domain within computer vision, with significant applications spanning security surveillance, autonomous navigation, and other fields. Throughout the tracking process, distractors and target appearance variations frequently arise, rendering sole reliance on initial templates unreliable. Therefore, the effective integration of spatiotemporal information and search region features plays a crucial role in achieving robust long-term single-object tracking. However, most existing methods indiscriminately incorporate all historical features as spatiotemporal context, potentially introducing irrelevant or redundant information that undermines tracking reliability. To address this limitation while more effectively exploiting backbone features, we propose FSTrack, which leverages feature fusion to enhance search features and adaptively selects features to strengthen spatiotemporal features. First, we integrate multi-level backbone features through feature fusion and enhance feature resolution, thereby fully exploiting the multi-scale features of the backbone networks. Second, we introduce an adaptive feature selection mechanism that dynamically identifies and emphasizes discriminative historical features, enhancing the robustness of spatiotemporal modeling under diverse tracking scenarios. Third, we propose a globally contextual prediction head that overcomes the limitation of the limited receptive field inherent in conventional CNN-based heads and further improving the overall performance. Extensive experiments demonstrate the superiority of FSTrack. On mainstream benchmark datasets such as GOT-10k, TrackingNet, and LaSOT, our approach outperforms mainstream models using both the same and higher resolution inputs in terms of speed and accuracy, achieving state-of-the-art results on tracking benchmarks.