Crowd Activity Change Point Detection in Videos via Graph Stream Mining.

In recent years, there has been a growing interest in detecting anomalous behavioral patterns in video. In this work, we address this task by proposing a novel activity change point detection method to identify crowd movement anomalies for video surveillance. In our proposed novel framework, a hyperspherical clustering algorithm is utilized for the automatic identification of interesting regions, then the density of pedestrian flows between every pair of interesting regions over consecutive time intervals is monitored and represented as a sequence of adjacency matrices where the direction and density of flows are captured through a directed graph. Finally, we use graph edit distance as well as a cumulative sum test to detect change points in the graph sequence. We conduct experiments on four real-world video datasets: Dublin, New Orleans, Abbey Road and MCG Datasets. We observe that our proposed approach achieves a high F-measure, i.e., in the range [0.7, 1], for these datasets. The evaluation reveals that our proposed method can successfully detect the change points in all datasets at both global and local levels. Our results also demonstrate the efficiency and effectiveness of our proposed algorithm for change point detection and segmentation tasks.