Combining OpenPose with BiLSTM for Violence Detection in Long-Term Care

Abstract

The Ministry of Health and Welfare's Statistics reports present the incidence of domestic care violence becomes higher annually. However, there is no efficient method to get rid of physical abuse. After being ill-treated of violence, someone is assessed injure by official organization. Then, the victims take legal actions to damage after the events. The deep learning motion recognized violence to family care in advance. To analyst that the images from complex data sets on the internet is important. The key part of images that are recognized as physical abuse is ambiguous and distorted in many pictures. The solution of ambiguity is to label joint points of human skeleton by OpenPose, and to train the marked joint point features in Bi-directional Long Short-Term Memory (BiLSTM). The accuracy is about to 96%, that can effectively detect physical abuse in time in the experimental results.