Robust learning for real-world anomalies in surveillance videos

Anomaly detection has significant importance for developing autonomous surveillance systems. Real-world anomalous events are far more complex and harder to capture due to diverse human behaviors and a wide range of anomaly types. A key factor in defining activity is the temporal length or duration of the activity. The time period required for an anomalous activity to be completely understandable and meaningful depends on the nature and speed of the event. Some events are as fast to be captured within a few frames; however, some activities are slow and may require several thousands of video frames to define an activity. Deep learning architectures have a limited input temporal sequence length and suffer from learning very long sequences. There is a need to re-investigate the problem from the frame sequences perspective to better define an activity in the limited temporal length. In this research work, our contribution is two-fold. Firstly, a novel strategy of dynamic frame-skipping is proposed for producing meaningful temporal sequences for model learning. Secondly, a new deep learning model based on the Inflated Inception network (I3D) is proposed for learning spatial and temporal information from video frames. In order to evaluate the performance of the proposed model, experiments are performed on one of the most challenging real-world anomalies UCF-Crime dataset. The results confirm that the proposed model is robust and significantly outperforms state-of-the-art methods in terms of accuracy. In addition to this, the proposed model has achieved the highest F1 score for fast and slow activities, such as explosions, road accidents, robbery, and stealing, and the AUC score of 0.837.