Detecting Anomalies in Videos Using Reconstruction and Prediction based Deep Learning Approach

Abstract

Anomaly detection in videos deals with pointing out the events that are out of normal. Current methods deals with identification of anomalous frames in a video sequence based on certain objects, and behaviours present in the video. Anomalies in videos are continuous events, and due to high number of features, generally the classical methods are not good enough for the task. Most of the reconstruction based deep learning methods works on the assumption that anomalies are rare in nature, and the training sets doesn’t contain any kind of anomalous events. This may work in case of object related anomalies, but will fail in case of motion related anomalies. We design a two-branch reconstruction and prediction based convolutional auto-encoder which utilises future frame prediction technique along with 3D convolutions to capture both spatial and temporal features. Moreover, the use of skip connections have been utilised in prediction branch to avoid the loss of spatial information during prediction in crowded frames. To overcome the problem of small dataset, we created new dataset by superimposing images over one another. This led to more data as well as frames containing more crowd density.