An Object-Aware Anomaly Detection and Localization in Surveillance Videos

Abnormal event detection plays an important role in video surveillance and smart camera systems. Existing methods in the literature are usually not object-aware, where different objects are not distinguished in processing. In this work, we propose an efficient object-aware anomaly detection scheme, specifically focusing on certain object categories, such as pedestrians. We first perform a block-based foreground segmentation to confine our analysis to moving objects and avoid irrelevant background dynamics. Then we discard uninterested objects by running an object detector on connected blocks. Finally we extract histograms of block-motion trajectories and cluster them to represent normal events. Our experiments demonstrate the accuracy and efficiency of the proposed method on dataset (PKU-SVD-B). We also propose a clip-based evaluation criterion with practical consideration and discuss this method at last.