Anomaly Detection in Automated Surveillance using Semi-Supervised Autoencoders

The recent large-scale deployment of cameras has paved way for a great demand to produce and use systems capable of capturing anomalous behaviors and events when provided with surveillance feeds. A great deal of tradeoff faced here is between managing the complexity of the algorithm and attaining high levels of accuracy and efficiency. This paper aims to achieve relatively high level of accuracy with less complexity on anomalous data by formulating a spatiotemporal autoencoder model and using ‘Semi-supervised’ learning approach. An autoencoder is a deep machine learning algorithm which encodes the input feature to a lower dimension and decodes the original input after reconstruction. It is usually unsupervised in nature. However, here semi-supervised learning technique has been used which aids in better clusterization of un-labelled data, increasing the accuracy of output. The proposed model has been tested with the benchmark i.e., UCF crime dataset. The proposed method has also been tested on other datasets, such as subway entrance, avenue and subway exit datasets to perform one-on-one comparison between different methods. Results obtained using proposed model have mostly improved compared to related research. The model was analyzed in both batch-wise and frame-wise modes. An anomalous event detection rate of 94% and 90% has been achieved for batch-wise and frame-wise, respectively. Computation time has been observed in batch mode. Though it has reduced mostly, however, it varies with length of video and type of crime.