Modelling A Framework To Obtain Violence Detection With Spatial-Temporal Action Localization

The system that detects and identifies human activities are named human action recognition. On the video approach, human activity is classified into four different categories, depending on the complexity of the steps and the number of body parts involved in the action, namely gestures, actions, interactions, and activities, which is challenging for video Human action recognition to capture valuable and discriminative features because of the human body’s variations. So, deep learning techniques have provided practical applications in multiple fields of signal processing, usually surpassing traditional signal processing on a large scale. Recently, several applications, namely surveillance, human-computer interaction, and video recovery based on its content, have studied violence’s detection and recognition. This paper has conducted an experiment to identify and detect violence with spatial-action localization, adapting a public dataset for effect. The idea was used an annotated dataset of general action recognition and adapted only for violence detection.