Vision-Based Vehicle Detection for VideoSAR Surveillance Using Low-Rank Plus Sparse Three-Term Decomposition
Ying Zhang, Daiyin Zhu, Peng Wang, Gong Zhang, Henry Leung
Abstract
Automatic vehicle detection from a video synthetic aperture radar (VideoSAR) system presents significant potential to enhance the surveillance performance in dynamic region of interest (DROI). In this paper, a novel VideoSAR low-rank plus sparse decomposition (LRSD) perspective for single-channel single-pass configuration is proposed to track the ground defocusing vehicles. Vehicle imaging features with 2-D motion parameters are derived theoretically by exploiting a priori knowledge of polar format algorithm (PFA). In accordance with the revealed characteristics, a vision-based VideoSAR-LRSD algorithm, called three-term decomposition (TTD) with proximal exchange-based alternating directions method of multipliers (PEADMM), is then proposed to improve the performance of vehicle detection. It can be used to break the limitation for the application of emergency response not permitting the acquisition of multi-channel or multi-pass data. We comprehensively demonstrate using extensive VideoSAR DROI experiments that in comparison with the state-of-the-art algorithms, TTD-PEADMM algorithm presents the improved accuracy and is able to offer competitive results.