Abstract: Aiming at the problem that the previous foreground detection methods depend more
heavily on scene information, a real-time foreground detection deep learning model ForegroundNet
without iteratively updating the background model is proposed. ForegroundNet extracts semantic
features from current and auxiliary images with backbone networks firstly, the auxiliary images which
can be either an adjacent image frame or an automatically generated background image. These
features are further fed into deconvolution network with short connections, which make the final
feature maps have the same size as input images and contain semantic and motional features in
different scales, finally we use softmax layer to perform a binary classification. The results on CDNet
dataset show that ForegroundNet achieves better F-Measure of 0.94 compare to the 0.82 of
suboptimal method. More over ForegroundNet has good real-time performance that its speed reaches
123 fps.

Key words: foreground detection, deep learning, computer vision, convolution neural network, motion segmentation