Publications

Abstract : Anomaly detection in surveillance videos is a challenging and demanding task. Autoencoders trained on segments of normal events are expected to give high reconstruction error for abnormal events than that for normal events. However, the assumption of autoencoders giving high reconstruction error is not always true in practice. Since the autoencoder sometimes offers better generalization, it also reconstructs abnormal events well, leading to slightly degraded performance for anomaly detection. Another issue is that the performance of real-time anomalous activity detection in surveillance videos still needs improvement. To address these issues, we propose an Object-centric and Memory-guided residual spatiotemporal autoencoder (OM-RSTAE) to detect video anomalies. The proposed technique achieved improved results over benchmark datasets, namely UCSD-Ped2, Avenue, ShanghaiTech and UCF-Crime datasets.