Abstract: A foreground detection method based on semantic segmentation and background reference was proposed to solve the problem of foreign objects detection in the space between platform screen doors and train doors in metro stations． This method used a depth neural network—DifferentNet to detect foreign objects region in images． Firstly， the background image and the image to be detected were obtained during a metro stop at the platform． The feature pyramid was obtained by extracting the feature information of the images through the encoding part of the network， and feature maps of the two images were merged by concatenation． Then the foreground heat map of the image to be detected was obtained by calculating the feature difference in the decoding part． Finally，the heat map was thresholded and filtered to get detection results． The results show that this method can achieve a high foreground IoU of 81. 2% and F1-score of 89. 5% ． Furthermore，it reached 30 fps in speed． The proposed method performed better than traditional methods and other image segmentation networks without background reference．

Key words: metro, foreign object detection, deep learning, semantic segmentation, foreground detection