Violent activity detection through surveillance camera using deep learning

Abstract

Surveillance camera systems have been implemented in most parts of the world to combat the rising rate of criminal activities. In the hopes of making public places safer for everyone, computer vision has also aided in making these systems more sophisticated but reliable and efficient. However, we are yet to make them better. While modern systems are able to record incidents, they often do not do so intelligently in order to make it easier for law reinforcements to respond quickly enough to aid victims or stop more crimes from occurring. Hence for our thesis project, we intend to use computer vision on a surveillance system so that it is able to identify crimes such as physical altercation, harassment, hijacking, snatching, etc. In this model, an action recognition system will be used, where we will be using extracted images from video feeds from multiple sources, and all those sources (cameras) will be centrally connected to a server. The server will be connected to databases containing information about violent activities. Based on the feeds, a signal will be sent to the respective system if a particular activity is detected. This system is mainly based on image processing concepts using different neural networks like MobileNet-V2, ResNet50, and LSTM to match live images with the existing trained system. This model will specifically use to detect criminal activities such as punching, kicking, slapping, and weapon violence, and all these pieces of information will be previously stored in the database. We have also implemented Grad-CAM in an effort to apply model explain ability.