An Unmanned Aerial Vehicle Video Object Tracking Algorithm Based on Siamese Attention Network

Abstract

Unmanned Aerial Vehicle (UAV) has been widely used in military and civilian fields, and object tracking is one of the critical technologies in UAV application. For addressing deformation, occlusion, small object, and other UAV object tracking problems, an UAV video object tracking algorithm based on Siamese Attention Network (SANet) is proposed in this paper. Initially, we designed a lightweight network as an extractor to extract features. After that, the attention mechanism module is constructed to screen out the feature map's semantic attributes, and the corresponding weights are re-assigned to different channels and spatial features. Finally, three Regional Proposal Networks (RPNs) are introduced to hierarchical fusion to obtain the tracking results. Our proposed algorithm in this paper has experimented on the UAV123 dataset and self-built dataset. The results show that the algorithm has a good tracking effect, the average accuracy is improved to 0.815, and the success rate is 0.619.