Simple Approach for Violence Detection in Real-Time Videos Using Pose Estimation With Azimuthal Displacement and Centroid Distance as Features

Abstract

Detecting violence in real time videos is not an easy task even for the most advanced deep learning architectures, considering the subtle details of human behavior that differentiate an ordinary from a violent action. Even with the advances of deep learning, human activity recognition(HAR) in videos can only be achieved at a huge computational cost, most of the time also requiring special hardware for reaching an acceptable accuracy. We present in this paper a novice method for violence detection, a sub-area of HAR, which outperforms in speed and accuracy the state of the art methods. Our method is based on features extracted from the Pose estimator method OpenPose. These features are then transformed into more representative elements in the context of violence detection, which are then submitted to a LSTM neural network to learn how to identify violence. This work was inspired by the violencedetector.org, the first open source project for violence detection in real time videos.