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.
