Memory Clustering Autoencoder Method for Human Action Anomaly Detection on Surveillance Camera Video

Unsupervised deep-learning methods with a deep clustering model are widely used to detect anomaly human actions obtained by surveillance camera due to their powerful image feature learning abilities. These methods aim to optimize a model through a clustering induction target to provide useful cluster assignment and usually use a forward propagation process: an autoencoder (AE) reconstructs an input sequence, a clustering model provides cluster allocation, and a scoring model evaluates distribution and provides scores for each sample. However, these methods have two main problems: one is that an AE is difficult to handle human posture when abnormal and normal actions occur in crowd scenes captured by a surveillance camera. The other is that network updating is interrupted by feature extraction and clustering in one epoch. To solve these problems, we design a deep memory clustering method based on graph convolution AE (MC-GCAE) to implement the real-time updating of pseudo-labels and network parameters. We also design a new loss function to express the similarity between the sample feature and the centroid feature in a memory storage module and to constrain the parameter update of a network. We evaluate the unsupervised method for three important and representative datasets mainly composed of surveillance videos and use the area under ROC curve (AUC) score as an experimental evaluation indicator.