TCAT: 제조 환경 내 인간-로봇 협업을 위한 시간 일관성 기반 적응형 객체 추적 기법
TCAT: A Temporal Consistency-based Adaptive Object Tracking Method for Human-Robot Collaboration in Manufacturing Environments
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This paper proposes an enhanced tracking method for stable object recognition in Human- Robot Collaboration (HRC) manufacturing environments. Existing methods often struggle with occlusions by workers’ hands, robotic interference, and dynamic pose variations. To address these challenges, we introduce Temporal Consistency-based Adaptive Tracking (TCAT), which leverages temporal consistency analysis and adaptive weighting strategies. TCAT quantifies temporal consistency by analyzing detection confidence scores across three consecutive frames and employs sigmoid-based adaptive weighting to balance spatial IoU and appearance feature similarities effectively. We validate TCAT on a custom dataset comprising 100,000 images of five vacuum cleaner components, integrating it with state-of-the-art object detectors such as YOLOv10-12 and RT-DETR. Experimental results demonstrate that TCAT significantly outperforms baseline methods, achieving a mAP@0.5 of 85.7% with YOLOv12—a 3.4 percentage-point improvement over ByteTrack—while maintaining a real-time processing speed of 15 FPS on an NVIDIA Jetson Orin Nano. The proposed method shows robust tracking performance and computational efficiency, which is suitable for practical HRC applications.
Abstract
This paper proposes an enhanced tracking method for stable object recognition in Human- Robot Collaboration (HRC) manufacturing environments. Existing methods often struggle with occlusions by workers’ hands, robotic interference, and dynamic pose variations. To address these challenges, we introduce Temporal Consistency-based Adaptive Tracking (TCAT), which leverages temporal consistency analysis and adaptive weighting strategies. TCAT quantifies temporal consistency by analyzing detection confidence scores across three consecutive frames and employs sigmoid-based adaptive weighting to balance spatial IoU and appearance feature similarities effectively. We validate TCAT on a custom dataset comprising 100,000 images of five vacuum cleaner components, integrating it with state-of-the-art object detectors such as YOLOv10-12 and RT-DETR. Experimental results demonstrate that TCAT significantly outperforms baseline methods, achieving a mAP@0.5 of 85.7% with YOLOv12—a 3.4 percentage-point improvement over ByteTrack—while maintaining a real-time processing speed of 15 FPS on an NVIDIA Jetson Orin Nano. The proposed method shows robust tracking performance and computational efficiency, which is suitable for practical HRC applications.