Detecting Video Anomalous Events with an Enhanced Abnormality Score

Detecting video anomalous events is vital for human monitoring. Anomalous events usually contain abnormal actions with exaggerated motion and little motion. We define the former and the latter as dynamic anomalies and static anomalies, respectively. We define the video data of events where a few persons perform diverse actions indoors as Indoor Event Data (IED). Many frame prediction approaches have succeeded in detecting dynamic anomalies. However, they are prone to overlooking static anomalies in IED. To solve this problem, we propose an Enhanced Abnormality Score (EAS), which is a combination of prediction, dynamic, appearance, and motion scores. To specifically target static anomalies, we calculate a score to evaluate the dynamic degrees of actions. We use an appearance score of a frame to detect static anomalies from appearance. This score is generated from a clustering-based distance of a pre-trained CNN feature. We also use a motion score based on flow reconstruction to balance the appearance score. We conduct extensive experiments on two datasets involving indoor human activities. Quantitative and qualitative experimental results show that our proposal achieves the best performance among its variants and the state-of-the-art methods.