이상 행동탐지기반의 자동 선별적 비식별화 연구

초록

Until now, in automatic de-identification, face recognition or object detection has been used to de-identify a frame. However, de-identification of motion units is required in anomal situations (disgust, violence, etc.), necessitating video editing. In this paper, we propose a method for de-identification of motion units, using you only live once (YOLOv5) for object detection and multiscale vision transformer (MViT) technology for motion recognition in the spatial domain, accompanied by transformer-based anomaly detection in the temporal domain, to analyze the start and end of the corresponding frame sections to be de-identified. In this experiment, we de-identified a large number of videos, achieving an average accuracy of 72.8% for five types of anomalous behavior (falling down, assault, sitting down, accident, and vandalism).

키워드