Towards Intelligent Vision Surveillance for Police Information Systems

Traditional surveillance systems detect suspicious activities with the assistance of human operators watching screens showing video streams of activities captured from different cameras, which often leads to fatigue and failure to identify the suspicious activities. This is an area of considerable interest for security agents like police carrying out surveillance operations. Besides the current literature effort, this research investigates how suspicious loitering in a location can be detected before the crime occurs. Since human behavior is dynamic, this research develops an intelligent vision framework based on an integrated convolutional neural network (CNN) adaptive to specific locations at a time rather than building a generalized model prone to errors. We demonstrate the efficiency of the proposed system with preliminary results obtained on real-life image frames captured from multiple cameras. This model outperforms the conventional approaches in terms of the detection of suspicious locations with an average Fl-score of 0.9666, a false positive rate of 0.0922, and an accuracy of 94.49%. The deployment of this new model can help to augment the work of police information systems.