Main-Auxiliary Aggregation Strategy for Video Anomaly Detection

Surveillance video anomaly detection is to detect unusual events in videos. Different anomaly events have different anomalous properties. A single detector only can detect anomalies precisely on the property it focuses on. This letter proposes a Main-Auxiliary Aggregation Strategy (MAAS) to aggregate multiple detectors focusing on different properties. The MAAS consists of two branches: a main-detector branch and an auxiliary-detectors branch. We select and assign a detector to the main-detector branch to provide the basic anomaly detection ability. We assign the remaining detectors to the auxiliary branch to provide high-precision decisions on other properties. In the auxiliary branch, we set strict thresholds adaptively for each detector to detect strong-abnormal (SA) and strong-normal (SN) samples as high-precision decisions. Then, we design a 1D-Floodfill method to deduce more high-precision decisions according to events’ temporal-continuity property. Finally, we design a soft-weight method to fuse the high-precision decisions in the auxiliary branch with the main detector to detect anomalies. We carry out experiments on multiple datasets. The experiments demonstrate the superiority of MAAS over the current state-of-the-art (SOTA) methods.