A Macro-observation Approach of Intelligence Video Surveillance for Real-Time Unusual Event Detection

In this paper, we propose a macro-observation scheme for unusual event detection in daily life, where motions in time-space domain are described by a global representation and individual activities do not have to be defined and modeled beforehand. The proposed representation records the time-space energy of motions of all moving objects in a scene without segmenting individual object parts or tracking objects. In daily life, images from a video sequence that spans sufficient repetition of normal day-to-day activities are first randomly sampled. A hierarchical fuzzy C-means clustering is used to divide the sampled images into groups. The goal of clustering is to assign similar training samples to the same cluster so that the distance of every member in the cluster to the cluster center meets a minimum distance threshold. The new observed event that has distinct distance from any of the cluster centroids is then classified as an anomaly. The proposed method has been evaluated in daily work of a laboratory and a convenience store. In order to test the robustness of the proposed method for unusual events detection in daily life, the laboratory scene was continuously monitored for 31 days. The experimental results reveal that the proposed method can well detect unusual events such as fighting as long as they last for a sufficient duration of time.