SiamTBD: A Novel Algorithm for Small and Dim Target Tracking Using Video SAR

A large number of publications on detection and tracking have appeared in the field of video synthetic aperture radar (Video SAR) due to its capability of forming high-resolution videos for a region of interest. However, Video SAR tracking faces many challenges, e.g., the dim and small target shadows with weak visual features, interference from nearby similar objects, etc. Moreover, the target may be submerged in strong clutters for a short period. In this work, a novel Siamese Region Proposal Network (SiamRPN) based track-before-detect algorithm, namely SiamTBD, is proposed to address the above issues. The SiamRPN network is chosen as the tracking backbone, considering its superiority in speed and accuracy. Moreover, a multi-frame tree composed of root/branch/leaf bounding boxes in different frames is constructed based on multi-frame information. In each frame, we select multiple templates based on the scores assigned to the branch/leaf bounding boxes, which are calculated based on both the appearance features and the motion information. Furthermore, inspired by the idea from the track-before-detect (TBD) technique, the information from multiple frames is accumulated and finally used to determine the existence of the target as well as estimate its trajectory. A number of experiments are conducted using the Video SAR data released by Sandia National Laboratory. The experimental results show that the proposed SiamTBD algorithm outperforms several state-of-the-art correlation filter-based and Siamese network-based tracking algorithms.