A Hog & Graph Based Human Segmentation from Video Sequences

Human segmentation from video is a significant problem to recognize a specific person which is desired to find. There are a remarkable number of studies discussing segmentation process in videos. Almost all study is examined how well a human segmentation process could be created by considering their position, clothing and face. Unfortunately, it has been assumed that the background is static by approximately all of these works. Main challenges in this research area are; non-static background derived from dynamic structure of videos, human body and face position, shadows and clothing changing after a period of time in videos. In this script, we have proposed a multi-model (Histogram of the Gradients - HOG and Graph-based) human segmentation technique that has worked with respect to HOG features and detected low-level and mid-level features of human clothing by supposing their positions in video. The offered technique is designed to demonstrate robustness against such challenges emphasized above. In this study, a well-known video series have been used. The video scenes can reach 15~25 fps and have the size about 640×480px. To compare graph-based method robustness a well-known segmentation method Watershed is also experimented and both methods are simply compared. Eventually, we determined that the proposed technique can produce satisfactory quality segmentation mentioned and detailed in following sections.