CAPS : CCTV 영상을 이용한 자율형 딥러닝 기반 아동학대 감지 시스템

Abstract

Though the mandatory policy of installing CCTV in the childhood care facilities of public institutions such as kindergarten and daycare center, the criminal of child abuse cases is gradually increasing due to the lack of awareness of violent acts and the difficulty in understanding the reporting processes. This paper proposes a novel Child Abuse Protection System (CAPS) to solve the above social problem. The proposed CAPS is composed of three functional software modules to implement a deep-learning-based system that autonomously detects violent acts against children. First, the clip creator module divides long CCTV videos into several pieces of short video clips. Second, the violence detector module classifies the abuse behaviors from the generated clips. Finally, the face detector module automatically processes the witnessed suspect’s face being blurred out by mosaic. Experimental evaluation results show that the most suitable feature extractor for detecting the child abuse behaviors is the MobileNetV2+LSTM model among several candidates of the proposed CNN+LSTM violence detection module, which has the best at 92.51% accuracy. Furthermore, the recall rate can be increased up to 6% by exploiting the proposed data augmentation technique. Codes are available at https://github.com/learningsteady0J0/ CAPS-Child-Abuse-Protection-System.