SSHA: Video Violence Recognition and Localization Using a Semi-Supervised Hard Attention Model

Current human-based surveillance systems are prone to inadequate availability and reliability. Artificial intelligence-based solutions are compelling, considering their reliability and precision in the face of an increasing adaption of surveillance systems. Exceedingly efficient and precise machine learning models are required to effectively utilize the extensive volume of high-definition surveillance imagery. This study focuses on improving the accuracy of the methods and models used in automated surveillance systems to recognize and localize human violence in video footage. The proposed model uses an I3D backbone pretrained on the Kinetics dataset and has achieved state-of-the-art accuracy of 90.4% and 98.7% on RWF and Hockey datasets, respectively. The semi-supervised hard attention mechanism has enabled the proposed method to fully capture the available information in a high-resolution video by processing the necessary video regions in great detail.