VLM 기반 장면 맥락 인지와 포즈 정렬 ROI를 결합한정책 중심 PPE 준수 판별 프레임워크

초록

Personal protective equipment compliance is essential for worker safety in industrial environments. However, most existing automated monitoring studies only determine whether PPE is worn, regardless of site-specific conditions. This paper proposes a policy-centric PPE compliance framework that infers required PPE from visual risk cues and verifies wearing evidence using pose-based body-part ROIs. The framework consists of pose estimation, ROI generation, VLM-based requirement inference, and compliance determination. Keypoint-based rotated ROIs improve part-level evidence collection in multi-worker environments. On the SH17 dataset, the proposed framework achieved 81.27% end-to-end exact-match accuracy, and upper-body ROI optimization improved safety-vest coverage recall from 11.95% to 70.83%.

키워드