Abstract: Abstract: In this paper， we propose a multi-scale deep autoencoder network based on dilated convolution， incorporating pedestrian prototypes and spatio-temporal features. To better exploit the temporal features of pedestrians in videos， a dual-branch structure is added to the potential space of the encoder and decoder， the ST-RNN branch of the recurrent neural network for predicting spatio-temporal features and the memory storage module for preserving the normal patterns of pedestrians. To enhance pedestrian feature extraction， ignore the effect of background information，and improve the generalization ability of the model， an improved atrous spatial pyramid pooling （ASPP） module is added in the encoder， the hybrid dilated convolution （HDC） principle is used in the convolution block to solve the pedestrian size variation problem， while a multi-level residual channel attention mechanism is introduced in the decoder to obtain more contextual information. The corresponding area under the ROC curve （AUC） of this model reaches 0.982， 0.928 for USCD ped2， CUHK Avenue datasets， respectively.

Key words: Key words: hybrid dilated convolution, residual attention, anomalous behaviour detection, deep convolutional autoencoder