Neural Network-Based Human Segmentation in Video Sequences Using Background Subtraction Map

Image segmentation is the task of assigning each pixel in an image the object class (one of the predefined) that it belongs to. Human instance segmentation from the background in video is a hot challenge for surveillance applications. In this paper, we examine if background subtraction —— a technique to extract all foreground objects in video from static camera —— can help to tackle this challenge. For this, we employ a recent successful deep learning method for still image segmentation and experiment with incorporating background subtraction in it. As a result, we propose an algorithm that yields fairly accurate and visually coherent results, and needs only a single additional hardware unit to run real-time. Finally, we elaborate on some important details inferred from the study as well as on future research directions.