본문 바로가기
  • Home

Advanced Strategy for On-Device Single Object Tracking Algorithm Robust to Multi-Domain Shift and Occlusion

  • Journal of The Korea Society of Computer and Information
  • Abbr : JKSCI
  • 2026, 31(7), pp.95~106
  • Publisher : The Korean Society Of Computer And Information
  • Research Area : Engineering > Computer Science
  • Received : April 30, 2026
  • Accepted : June 17, 2026
  • Published : July 31, 2026

Ji-Su Yoon 1 Jae-Yeob Kim 1 Hyun-Sung Jang 1 Seung-Wook Park 1

1엘아이지 디펜스앤에어로스페이스

Accredited

ABSTRACT

In this paper, we propose a robust object tracking tuning strategy suitable for gimbal control in mobile platforms, ensuring stable operation across diverse scenario. The core method applies domain generalization for robustness against unseen environments and introduces an occlusion-aware head, trained with spatial Cox process-based random masking and pseudo-labels, to autonomously recognize occlusion states. Additionally, to prevent bounding box jittering during tracking failures, we employ a rule-based post-processing algorithm that detects outliers in the bounding box center shift and real-time response map statistics using a Z-score approach. This immediately corrects trajectory instability, maintaining stable outputs for gimbal systems. Experimental results show that the proposed model achieves superior robustness over baselines. Furthermore, it operates at 77.5 FPS in a high-performance GPU environment and 30.3 FPS in an edge device environment, proving its high efficiency in satisfying both performance and real-time requirements in resource-constrained on-device environments.

Citation status

* References for papers published after 2025 are currently being built.