End-to-End Video Surveillance Framework for Anomaly Detection and Person Re-identification

Unmanned surveillance systems represent a cutting-edge frontier in security technology, offering enhanced monitoring capabilities while minimizing reliance on constant human oversight. However, traditional approaches suffer from limitations due to manual monitoring, leading to potential lapses in critical event detection. In response, our project introduces a unique semi-supervised approach using the UCF-Crime dataset [1] to automate surveillance by integrating the state-of-the-art MGFN [2] model with our custom LSTM multi-class classifier and re-identification model. While MGFN [2] provides binary classification with an AUC of 86.98%, our implementation uses this along with our custom multi-class classifier which has an AUC of 83%, to predict specific categories of anomalies like burglary, abuse, and fighting. The detected offenders are noted and are looked for in every other feed of video as and when they appear using DeepSORT [3] re-identification model. Moreover, our system notifies authorities about anomalies and identifies individuals through a comprehensive dashboard interface. This fusion of models allows for more nuanced event detection, contributing to the advancement of surveillance technology and public safety.