Reverse erasure guided spatio-temporal autoencoder with compact feature representation for video anomaly detection

Yuanhong Zhong, Xia Chen, Jinyang Jiang, Fan Ren

Abstract: In conventional video anomaly detection based on deep learning, the deep network is optimized without focus and the similarity between different normal frames is ignored. To alleviate these issues, we designed a dualencoder single-decoder network to reconstruct frames and proposed a training strategy involving reverse erasure based on the reconstruction error and deep SVDD to regularize the training of the network. With this training strategy, the proposed model achieved high performance in terms of both the AUC and EER. Future work will involve the application of our training strategy to more complex tasks.