Are there any papers implementing multiple people tracking algorithms (primarily the silhouette part) using depth data from Kinect (or Asus Xtion)? RRS feed

  • Question

  • EDIT: A same post can be seen on

    Any response would be apreciated~ 


    I'm trying to implement the Skeletoning algorithm which is already implemented in Kinect and OpenNI-NITE, to see if there could be any further improvements.

    As I know, both the Kinect and NITE's implementations on Silhouetting are quite good at handling partial overlapping conditions and other complicated cases, yet they both are patented, so I'd like to know if there are any papers talking about Silhouetting that Kinect (or NITE) based on?

    While I'm searching through Google Scholar, I find it's hard to find a paper directly related to "silhouette" algorithm. Most of them are of topics "background subtraction", "human detection", "people tracking", which care neither about how many people are in the scene, nor which person exactly does one pixel belongs to. Here are some papers that I'm reading:
    [1]. Xia L, Chen C C, Aggarwal J K. Human detection using depth information by Kinect[C]//Computer Vision and Pattern Recognition Workshops (CVPRW), 2011 IEEE Computer Society Conference on. IEEE, 2011: 15-22.
    (It relies on 2D & 3D template matching of human heads, which is actually quite unstable when head is not well matched or even not in the scene)

    [2]. del-Blanco C R, Mantecón T, Camplani M, et al. Foreground Segmentation in Depth Imagery Using Depth and Spatial Dynamic Models for Video Surveillance Applications[J]. Sensors, 2014, 14(2): 1961-1987.
    (It's roughly an approach of probablity, which will lose the human silhouette when he stops moving, and it doesn't care about the case that multiple people overlapping partly, and so does the following link [3])

    [3]. Wang C, Liu H, Ma L. Depth Motion Detection--A Novel RS-Trigger Temporal Logic based Method[J]. Signal Processing Letters, IEEE, 2014, 21(6): 717-721.

    [4]. Störmer A, Hofmann M, Rigoll G. Depth gradient based segmentation of overlapping foreground objects in range images[C]//Information Fusion (FUSION), 2010 13th Conference on. IEEE, 2010: 1-4.
    (This paper handles multiple overlapping people, yet with poor performance)

    Sunday, August 16, 2015 6:00 AM

All replies