Anomaly detection in surveillance videos using deep autoencoder

Video anomaly detection algorithms are yet to advance at the pace CCTV footage data of public places is being recorded and made publicly available. An anomaly specifies unusual activity or response in a video by one or more subjects/objects present in the video clip. Autoencoder being a powerful unsupervised method, has been popularly used for anomaly detection in various domains including video. In this work, we present a novel anomaly detection algorithm using deep autoencoders which exploits spatiotemporal features of training video clips along with a novel combined regularity score-based thresholding mechanism. Our model achieved AUC of 86.4% and 88.9% respectively over UCSD Peds1 and Avenue datasets successfully which is comparable with the existing works on video anomaly detection using autoencoders.