Anomaly Detection in Surveillance Video using Motion-Direction Model

In this paper, we proposed an anomaly detection technique based on Motion and Direction called MD that detect anomaly in crowded scenes where the traditional tracking approaches tends to fail. MD follows supervised learning approach that requires training of the model with labeled data. Initially input frames are divided into cells of size 10x10x5 pixels and then subjected to foreground segmentation that confines analysis to foreground pixels only. Low level features such as motion and direction are extracted and analyzed independently to detect anomalies in surveillance videos. We compared our proposed method MD with the recent approaches like SF, MPPCA, SF-MPPCA and MDT and demonstrate that we obtain better results than these methods except MDT. MDT outperforms all the above mentioned methods including our proposed method MD, but it requires more computation time to analyze entire frame and it is difficult to infer the nature of anomaly as it is a joint model of motion and appearance. Whereas, independent analysis by the proposed method MD infers the nature of anomaly and keeps the computation faster as cells are of very small size compared to frames.