부산대학교 도서관

Visual object tracking has been intensively studied for its role in traffic surveillance, human action recognition, and autonomous driving. Siamese network-based methods have demonstrated a satisfactory trade-off between precision and efficiency for visual tracking. Nevertheless, the accuracy of Siamese trackers is limited when it comes to predicting the target’s location in scenarios involving background clutter, changes in illumination, variations in scale, deformation, fast motion, among others. We suggest a novel approach in our manuscript, which involves a deep adaptive discriminative Siamese network equipped with an advanced fusion scheme for multiple level responses. To enhance the feature discriminability of the Siamese network, we introduce a novel residual channel attention clipping unit. This unit seamlessly integrates residual connections and channel attention, leading to significant optimization and improved representation in the network. Then, we introduce a multi-response adaptive fusion structure that takes the advantages of the low-level, mediate-level, and high-level features, yielding a comprehensive score map that reveals multiple levels of semantics. Our experiments demonstrate that our tracker performs exceptionally well compared to current leading trackers on widely-used public tracking datasets such as OTB-2015 and GOT10k. The method attains an AUC score of 0.655 on OTB2015, while maintaining a processing speed of 63 FPS.