상용 스마트글래스-스마트폰 연동형 단안 RGB 기반 근접 위험 인지 시스템

초록

Commercial smart glasses provide first-person RGB video; however, they generally lack depth sensors and sufficient on-device computing resources. This study presents a mobile system that analyzes monocular video from Meta Ray-Ban Display on a Galaxy S24+ to identify nearby hazards. Depth Anything V2 (DAv2) estimates relative depth, whereas SegFormer excludes low-risk regions, and MediaPipe removes the user’s hands. The remaining regions are scored based on proximity, size, and shape to generate CLEAR, CHECK, or OBSTACLE states. In 150 trial-level tests, the system achieved 95.33% accuracy and a 96.52% F1-score. In repeated 30-second video tests, its frame-level accuracy was 96.41%, compared with 54.38% for YOLOv8n. The corresponding wall-and-corner rates were 99.52% and 1.28%. The proposed method processed 2.46 fps, with mean display and vibration latencies of 516.25 and 534.45 ms. Removing SegFormer increased the empty-corridor false-alarm rate from 9.14% to 99.62%, confirming the importance of semantic exclusion.

키워드