어텐션 모듈을 이용한 딥러닝 기반의 폭력 탐지

Abstract

To prevent violent crimes, surveillance cameras have been deployed in public places. But, it is time and labor consuming to manually monitor a large amount of video data from surveillance cameras. Therefore, automatically detecting violent behaviors from video is essential. Existing methods tend to misclassify moving objects as violence. In order to improve this drawback, we propose to use spatial and channel features more efficiently using attention modules. The proposed method is based on the Flow Gated Network, 3D convolution layer and CBAM module. Experimental results have shown the proposed method achieves 1% improvement in accuracy, compared to the existing method.