Video Anomaly Detection

The rapid evolution of society has made the need for mass video surveillance in shops, parks, societies, and streets an imperative. This has escalated the need for effective video anomaly detection software that can accurately detect anomalous activities in varied contexts. Current methods of detection are largely reliant on contextual data and cannot recognise anomalies across a broad spectrum. This study proposes a multiple-instance based learning-based model that can detect anomalies in any context. The ranking nature of the model helps frame anomaly detection as a regression problem and arrives at a threshold level for anomalous frames. It use a reconstruction based approach to generate anomaly scores based on a batch of initial frames of the video to decide if a given frame is anomalous or isn’t. This study also introduces a new dataset of Indian-specific actors recreating criminal activities to provide racial breadth to the current datasets, which are mostly composed of Caucasian actors.