Anomaly Detection using a Convolutional Winner-Take-All Autoencoder

We propose a method for video anomaly detection using a winner-take-all convolutional autoencoder that has recently been shown to give competitive results in learning for classification task. The method builds on state of the art approaches to anomaly detection using a convolutional autoencoder and a one-class SVM to build a model of normality. The key novelties are (1) using the motion-feature encoding extracted from a convolutional autoencoder as input to a one-class SVM rather than exploiting reconstruction error of the convolutional autoencoder, and (2) introducing a spatial winner-take-all step after the final encoding layer during training to introduce a high degree of sparsity. We demonstrate an improvement in performance over the state of the art on UCSD and Avenue (CUHK) datasets.