Region-Based Anomaly Localisation in Crowded Scenes via Trajectory Analysis and Path Prediction

In this paper, we propose an approach for locating anomalies in crowded scene for surveillance videos. In contrast to the previous approaches, the proposed approach does not rely on traditional tracking techniques which tend to fail in crowed scenes. Instead the anomalies are tracked based on the information taken from a set of anomaly classifiers. To this end, each video frame is divided into non- overlapping regions wherein a set of low-level features are extracted. After that, we apply the anomaly classifiers which determine whether there is anomaly in each region. We then derive the anomaly trajectory by connecting the anomalous regions temporarily across the video frames. Finally, we propose path prediction using linear Support Vector Machine (SVM) to smooth the trajectory. By doing this, we will able to better locate them in the crowded scene. We tested our approach on UCSD Anomaly Detection dataset which contains crowded scenes and achieved notable improvement over the state-of-the-art results without sacrificing computational simplicity.