LLM 기반 시맨틱 스타일 분석을 활용한 사용자 추적 시스템
User Identification and Tracking System via LLM-Based Semantic Style Analysis
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This paper proposes a user tracking system based on semantic style recognition using a Large Language Model (LLM). To overcome the limitations of traditional face recognition in non-frontal or occluded scenarios, the proposed system extracts clothing styles and colors using GPT-4o, a multimodal LLM, and combines this information with YOLOv5-based segmentation and DeepSORT tracking. The extracted style features serve as robust identity cues that enable consistent tracking even when a user’s face is not visible. Experiments conducted on RGB-D image sequences with three subjects demonstrate the system’s high recognition accuracy and improved ID stability compared to conventional tracking methods. The integration of LLM-based style analysis enables ID restoration after tracking failures and enhances tracking performance in complex environments. The proposed framework shows strong potential for use in indoor robotics, smart surveillance, and human-robot interaction applications.
Abstract
This paper proposes a user tracking system based on semantic style recognition using a Large Language Model (LLM). To overcome the limitations of traditional face recognition in non-frontal or occluded scenarios, the proposed system extracts clothing styles and colors using GPT-4o, a multimodal LLM, and combines this information with YOLOv5-based segmentation and DeepSORT tracking. The extracted style features serve as robust identity cues that enable consistent tracking even when a user’s face is not visible. Experiments conducted on RGB-D image sequences with three subjects demonstrate the system’s high recognition accuracy and improved ID stability compared to conventional tracking methods. The integration of LLM-based style analysis enables ID restoration after tracking failures and enhances tracking performance in complex environments. The proposed framework shows strong potential for use in indoor robotics, smart surveillance, and human-robot interaction applications.