A multi-stage fusion instance learning method for anomalous event detection in videos

Jian Cheng, Fengquan Zhang, Guiling Wang, Wancai Zhang

Abstract: Anomalous event detection in giant amount surveillance footage in real world is currently an active research area. Variery and rareness of the anomaly events is still a thorny challenge to deal with. In this paper, we propose a multi-stage fusion instance learning method (MFIL) for inferring anomalous event pattern and predicting anomaly appearance in videos. We propose object-aware model and action-aware model to represent regularities of human objects and actions among frames exploiting cascaded deep network models. Furthermore we improve and represent fusion instance learning method for fetching and maximizing anomaly scores via object and action regularities in anomalous sequences from videos. We validate the performance of MFIL on action movie and UCF-Crime respectively, both contain anomalous and violent events. Experimental results demonstrated that MFIL is effective for anomalous event detection in videos gathered from real world.

External IDs:dblp:journals/mlc/ChengZWZ23