A Computer Vision Model for Detecting Suspicious Behaviour from Multiple Cameras in Crime Hotspots Using Convolutional Neural Networks

Abstract

Conventional security surveillance systems detect suspicious behaviours by the active participation of human operators constantly watching monitors showing video streams of activities captured from different cameras. The datasets are manually retrieved and analyzed after the occurrence of the incidents. This often leads to misinterpretation and late detection in a real-life environment. Recent studies have investigated the current surveillance systems for the challenges of the detection of suspicious incidents. The research question here is: How can suspicious breaking into run be detected being a common red flag leading to most crimes? This research develops a computer vision framework based on feature engineering, convolutional neural network (CNN), and median filtering to address the lacuna faced by the surveillance systems. Experiments were conducted on real-life image frames captured from multiple camera datasets. The proposed model outperforms the conventional approaches in terms of the detection of suspicious behavioural patterns with an average F1-score of 0.9661, a false positive rate of 0.0734, and an accuracy of 94.81%. The deployment of this proposed model in a crowded environment can help to augment the work of security personnel in raising awareness regarding possible crime at hot spots.