Real-time human tracking by detection based on HOG and particle filter

In this paper, a novel approach is proposed to achieve the multi-human tracking in video surveillance system using a combination of tracking by detection method. First of all, a modified foreground objects detection method is applied to extract the foreground blobs thus achieving the regions of interest. On the second hand, the HOG features together with searching strategies are used to initialize the trackers of humans and the trackers are utilized for the particle filter tracking. Through the tracker initialization, the problems that one blob might contain several people could be overcome. Moreover, for the tracking aspect, we further utilize the data from the foreground detection that a color-edgetexture histogram is used by calculating the local binary pattern of the edge of the foreground objects which could have a good performance in describing the shape and texture of the objects. Finally, occlusion solutions strategies are applied in order to overcome the occlusion problems during tracking. Experimental results on different data sets have proved that our method has better performance and good real-time ability.