An autoencoder is trained to generate the background from the surveillance image by setting the training label as the shuffled input, instead of the input itself in a traditional autoencoder. Then the multi-scale features are extracted by a sparse autoencoder from the surveillance image and the corresponding background to detect foreground.