Anomaly Detection using Convolutional Spatiotemporal Autoencoder

In this paper an efficient technique for detecting anomalies in videos is proposed. Most of the applications of Convolutional Neural Networks (CNNs) are based on object detection and recognition. Generally Convolutional layers are used in images, however they are supervised and need labels for learning which is cumbersome task. Proposed method includes spatio-temporal model for detection of anomalies in videos of crowded places like walkways, bus or railway stations etc. The model consists of two significant aspects, one for representation of spatial features and one for temporal progression of spatial features. Convolutional layers are used for spatial feature representation and Long Short Term Memory network (LSTM) is used for temporal progression representation. Anomaly score is introduced to detect the event is anomalous or normal. Experimental accuracy is comparable to state-of-the-art methods on standard databases like Avenue and UCSD.