Explainable Anomaly Detection in Videos Based on the Description of Atomic Actions

Abstract

Anomaly detection in surveillance videos is an exhaustive and tedious task to be performed manually by humans. Many methods have been proposed to detect anomalous events by learning normal patterns and differentiate them from abnormal ones. However, these methods often suffer from false alarms, as human behaviors and environments can change over time. In addition, these methods fail to discriminate the types of anomalies that can occur, especially in anomalies performed by humans. This work presents an approach to detect anomalous events based on atomic action descriptions. It combines a tracking people method with atomic action detection and recognition network to understand video events and generate atomic descriptions. Besides detecting the anomalies, the proposed approach can also describe the anomalous action with human attributes in natural language. Anomalies are detected based on the generated descriptions of the scene. Experimental results show the effectiveness of our approach, presenting an average F1-Score of 87%.