Background-aware Siamese Network Tracking Based on Salient Feature Fusion
Lifang Zhou,Jing Dai,Bangjun Lei,Rong Xue
DOI: https://doi.org/10.21203/rs.3.rs-3862057/v1
2024-01-01
Abstract:In the realm of object tracking, complex backgrounds pose significant challenges, often leading to the underperformance of existing trackers that predominantly rely on deep features from the final layer of feature extraction networks. These features, while semantically rich, are not always sufficient to distinguish targets from cluttered environments. To address this limitation, we introduce the Background-aware Siamese Network (BASNet), an innovative approach that enhances the salience of features in intricate scenes. Central to our method is the Dual-Feature Fusion (DFF) model, meticulously crafted to stabilize target representation amidst distracting backgrounds. BASNet leverages the complementary strengths of deep and shallow features, harnessing the former’s semantic depth and the latter’s precision in spatial localization. This fusion not only elevates feature utilization but also mitigates the shortcomings associated with the superficial output of shallow networks. Further refining the target’s spatial definition, our attention-focusing module (AF) plays a pivotal role. It accentuates pertinent features 1 and attenuates noise emanating from the multi-layer feature amalgamation. The module’s ingenuity lies in its cyclic attention operations, which enable each pixel to establish extensive dependencies with all others , thereby spotlighting salient features. For precise target localization, BASNet employs the SIoU loss within its regression branch, a strategic choice that contributes to the model’s accuracy. Extensive experiments demonstrate the competitive performance of our method on six datasets: OTB50, OTB100, UAV123, VOT2016, VOT2018, and VOT2019.