文献类型：期刊文献

英文题名：Target signature network for small object tracking

作者：Liang, Lei[1];Chen, Zhihua[1];Dai, Lei[1];Wang, Shouli[1]

机构：[1]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China

年份：2024

卷号：138

外文期刊名：ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE

收录：;EI(收录号：20244217219885);WOS:【SCI-EXPANDED(收录号:WOS:001338859700001)】；

基金：This work was supported by the National Natural Science Founda-tion of China (Grant No. 62272164 and No. 62306113) , and the Shang-hai Scientific and Technological Innovation Action Plans-Scientific Instrument Development, China (Grant No. 21142201300) .

语种：英文

外文关键词：Siamese network; Small object tracking; Size-aware information; Center attention; Dynamic positive samples

摘要：The wide application of drones has facilitated the advancement of small object tracking. However, limited features and background distractions pose challenges to modern trackers. Although progress has been made by learning more discriminative features, they still have difficulty under complex scenes. In this paper, a target signature network is tailored with awareness to size changes and robustness to background disturbances, where the target is identified implicitly with a signature composed by its size, center, and selection region of positive samples. For enriching size information, a size-aware module aggregates features through cascaded receptive field enlargement blocks in a multi-receptive learning manner. For filtering irrelevant noises, a center attention module is put forward to locate the target robustly. Moreover, we design a dynamic positive sample definition strategy to introduce more samples as the target gets smaller during training, thereby alleviating tracking drift or loss when its scale changes. Extensive comparisons with leading trackers and thorough ablation studies are conducted. On five widely adopted benchmarks, the proposed method surpasses the latest best-performing tracker by 2.2% and 1.2% for tracking success rate and precision on average. Superior performance has underscored its high efficacy.