Human segmentation in surveillance video with deep learning

Advanced intelligent surveillance systems are able to automatically analyze video of surveillance data without human intervention. These systems allow high accuracy of human activity recognition and then a high-level activity evaluation. To provide such features, an intelligent surveillance system requires a background subtraction scheme for human segmentation that captures a sequence of images containing moving humans from the reference background image. This paper proposes an alternative approach for human segmentation in videos through the use of a deep convolutional neural network. Two specific datasets were created to train our network, using the shapes of 35 different moving actors arranged on background images related to the area where the camera is located, allowing the network to take advantage of the entire site chosen for video surveillance. To assess the proposed approach, we compare our results with an Adobe Photoshop tool called Select Subject, the conditional generative adversarial network Pix2Pix, and the fully-convolutional model for real-time instance segmentation Yolact. The results show that the main benefit of our method is the possibility to automatically recognize and segment people in videos without constraints on camera and people movements in the scene (Video, code and datasets are available at http://graphics.unibas.it/www/HumanSegmentation/index.md.html).