Video anomaly detection and localization based on appearance and motion models

In this paper, we present an approach to detect and localize anomalies in the surveillance videos. Precise detection, modeling the normality in a context and dealing with false alarms are the major challenges to cope with while performing detection and localization of anomalies. We propose appearance and motion models to detect an anomalous event at the frame level in a video. In the anomalous frames, we localize the anomaly using appearance and connected component analysis cues. The proposed appearance model utilizes the mask-rcnn for scene semantic analysis. The proposed motion model constructs frame descriptors based on the histograms of the intensity difference maps of the consecutive video frames. A one class support vector machine (OCSVM) is used to detect motion based anomalous event. In order to eliminate the false alarms of anomaly detection and localization emerging due to camera jitter and object movements in unlikely motion regions of a video frame, we exploit the connected component analysis. Experiments on the Avenue and UMN datasets reveal that the proposed approach achieves high accuracy and outperforms the existing methods.