Grey Relational Frame Prediction Method for Anomaly Detection

Anomaly detection is a common problem in security and protection systems. Unlike conventional methods, considering the feature correlation and the uncertain predicted frames, a grey relational frame prediction method is proposed for the anomaly detection task. The future frame prediction network is designed by adversarial learning, consisting of generative and discriminant modules. In order to solve the lack of feature correlation, we integrate Deng's grey relation into the generative module to calculate the correlation between the predicted features and previous features during training. Furthermore, the grey absolute relation is introduced to deal with the uncertainty of predicted future frames. This network is optimized with different loss functions that combine the adversary, grey relation, pixel, gradient, and optical flow. These losses can well measure the difference between the predicted future frames and real future frames in temporal, spatial, and feature aspects. Experiential results show the proposed method obtains the averaged AUC of 84.1%, 95.7%, 85.6% on UCSD Ped1, Ped2, and CUHK Avenue datasets, which are 1%, 0.3%, 0.5% higher than the network without grey relation analysis. Extensive experiments demonstrate the superiority of our model in anomaly detection.