Graph Stream Mining Based Anomalous Event Analysis

A major challenge in video surveillance is how to accurately detect anomalous behavioral patterns that may indicate public safety incidents. In this work, we address this challenge by proposing a novel architecture to translate the crowd status problem in videos into a graph stream analysis task. In particular, we integrate crowd density monitoring and graph stream mining to identify anomalous crowd behavior events. A real-time tracking algorithm is proposed for automatic identification of key regions in a scene, and at the same time, the pedestrian flow density between each pair of key regions is inferred over consecutive time intervals. These key regions are represented as the nodes of a graph, and the directional pedestrian density flow between regions is used as the edge weights in the graph. We then use Graph Edit Distance as the basis for a graph stream analysis approach, to detect time intervals of anomalous flow activity and to highlight the anomalous regions according to the heaviest subgraph. Based on the experimental evaluation on four real-world datasets and a benchmark dataset (UCSD), we observe that our proposed method achieves a high cross correlation coefficient (approximately 0.8) for all four real-world datasets, and 82% AUC with 28% EER for the UCSD datasets. Further, they all provide easily interpretable summaries of events using the heaviest subgraphs.