On-the-fly global activity prediction and anomaly detection

We propose a unified framework using Latent Dirichlet Allocation (LDA) for discovering behaviour global correlations over a distributed camera network. We explore LDA for categorising object motion patterns as local behaviours in each camera view before correlating these local behaviours globally over different physical locations in multi-camera views. In particular, a Temporal Order Sensitive LDA (TOS-LDA) is formulated to discover behaviour global temporal correlations of different durations among all camera views simultaneously. In addition, a novel on-line global activity prediction method is proposed based on which global anomalies can be detected on the fly. We validate the effectiveness of our approach using public multi-camera CCTV footages.