Deep Crowd Anomaly Detection by Fusing Reconstruction and Prediction Networks

Abnormal event detection is one of the most challenging tasks in computer vision. Many existing deep anomaly models are based on reconstruction errors, where training phase performed using only videos normal events and model then capable to estimate frame-level scores for an unknown input. It assumed that error gap between frames abnormal high during testing phase. Yet, this assumption may not always hold due superior capacity generalization neural networks. In paper, we design a generalized framework (rpNet) proposing series by fusing several options network (rNet) prediction (pNet) detect efficiently. rNet, either convolutional autoencoder (ConvAE) or skip connected ConvAE (AEc) can be used, whereas pNet, traditional U-Net, non-local block attention U-Net (aUnet) applied. The fusion both rNet pNet increases gap. Our have distinct degree feature extraction capabilities. One our (AEcaUnet) consists AEc with proposed aUnet has capability confirm better extract quality features needed video detection. Experimental results UCSD-Ped1, UCSD-Ped2, CUHK-Avenue, ShanghaiTech-Campus, UMN datasets rigorous statistical analysis show effectiveness models.