HRMOT: two-step association based multi-object tracking in satellite videos enhanced by high-resolution feature fusion.

Wu, Yuqi; Zhang, Xiaowen; Liu, Qiaoyuan; Xue, Donglin; Sun, Haijiang; Ren, Jinchang

Abstract

Multi-object tracking in satellite videos (SV-MOT) is one of the most challenging tasks in remote sensing, its difficulty mainly comes from the low spatial resolution, small target and extremely complex background. The widely studied multi-object tracking (MOT) approaches for general images can hardly be directly introduced to the remote sensing scenarios. The main reason can be attributed to: 1) the existing MOT approaches would cause a significant missed detection of the small targets in satellite videos; 2) it is difficult for the general MOT approaches to generate complete trajectories in complex satellite scenarios. To address these problems, a novel SV-MOT approach enhanced by high-resolution feature fusion (HRMOT) is proposed. It is comprised of a high-resolution detection network and a two-step based association strategy. In the high-resolution detection network, a high-resolution feature fusion module is designed to assist the detection by maintaining small object features in forward propagation. Based on high-quality detection, the densely-packed weak objects can be effectively tracked by associating almost every detection box instead of only the high score ones. Comprehensive experiment results on the representative satellite video datasets (VISO) demonstrate that the proposed HRMOT can achieve a competitive performance on the tracking accuracy and the frequency of ID conversion with the state-of-the-art (SOTA) methods.

Citation

WU, Y., ZHANG, X., LIU, Q., XUE, D., SUN, H. and REN, J. 2024. HRMOT: two-step association based multi-object tracking in satellite videos enhanced by high-resolution feature fusion. In Ren, J., Hussain, A., Liao, I.Y. et al. (eds.) Advances in brain inspired cognitive systems: proceedings of the 13th International conference on Brain-inspired cognitive systems 2023 (BICS 2023), 5-6 August 2023, Kuala Lumpur, Malaysia. Lecture notes in computer sciences, 14374. Cham: Springer [online], pages 251-263. Available from: https://doi.org/10.1007/978-981-97-1417-9_24

• Presentation Conference Type: Conference Paper (published)
• Conference Name: 13th International conference on Brain-inspired cognitive systems 2023 (BICS 2023)
• Start Date: Aug 5, 2023
• End Date: Aug 6, 2023
• Acceptance Date: Jul 28, 2023
• Online Publication Date: May 22, 2024
• Publication Date: Dec 31, 2024
• Deposit Date: Jun 14, 2024
• Publicly Available Date: May 23, 2025
• Publisher: Springer
• Peer Reviewed: Peer Reviewed
• Pages: 251-263
• Series Title: Lecture notes in computer science (LNCS)
• Series Number: 14374
• Series ISSN: 0302-9743; 1611-3349
• Book Title: Advances in brain inspired cognitive systems
• ISBN: 9789819714162
• DOI: https://doi.org/10.1007/978-981-97-1417-9_24
• Keywords: Data association; High-resolution feature fusion; Multi-object tracking; Satellite video
• Public URL: https://rgu-repository.worktribe.com/output/2373020