Violence Activity Recognition Using Computer Vision

Abstract

I proposed an intelligent system algorithm to address real-time violence activity using computer vision. Sometimes in our absent different violence activity occurs in our daily life. As a part of a smart surveillance system detecting real-time violent activity plays a key role. A video is several frames of the pixel so analyzing and classify them is a challenging research topic in the field of computer vision. Deep learning nevertheless CNN is the key part of computer vision. In previous research action recognition mostly focus on real-life activities but not enough for predicting violence. Considering all possible situation to recognize real-life violence more accurately in this research I follow Convolutional Long Short-Term Memory (CONVLSTM). The model finds spatial features from video and analysis the correlation. Datasets collected from various source and comparatively I get an adequate accuracy result. The research project finished with several experiments using different deep video analyzing algorithms. I compared and differentiated different deep learning model and finalize the best one which about 90% accuracy result. Finally, realtime video footage set to classify with my trained model. The model returns the relevant output whether the scenario is violent or not at the same time the result sent through the cloud to my developed mobile application for further action.