Deep learning techniques for physical abuse detection

Abstract

Physical abuse has become a societal problem. Mostly children, women and old age people are vulnerable to it especially in cases of domestic violence or workplace aggression. Reporting it is in itself a challenge especially if there is a pre-existing relationship between the abuser and victim. In this paper we propose a deep learning technique for human action recognition and human pose identification to tackle physical abuse by detecting it in real time. 3D convolution neural network (CNN) architecture is built using 3D convolution feature extractors which extract both temporal and spatial data in the video. With multiple convolution layer and subsampling layer, the input video has been converted into feature vector. Human pose estimation is done using the detection of key points on the body. Using these points and tracking them from one frame to another gives spatial-temporal features to feed into neural network (NN). We present metrics to measure the accuracies of such systems where real time reporting and fault tolerance capabilities are of utmost importance. Weighted metrics shows accuracy of about 89.42% with precision of about 85.82% and thus shows the effectiveness of the system.