和歌山大学 システム工学部
Vision and Robotics Laboratory

  トップページ

  研究内容

  研究成果
    ”らしさ”に基づく対象検出・追跡
    ステレオビジョンシステム
    K-means Tracker (英語)
    Air Papyrus:3次元空間描画

  研究機材

  論文・発表

  スタッフ

  研究とは



  Private















K-means tracker:

A pixel-wise tracking algorithm with target and background samples

K-means tracker is a general tracking algorithm which achieves the robust object tracking under less constrained background conditions. By applying the K-means clustering to the selected target as well as background samples in a constructed 5D feature space, this algorithm can remove the mixed background pixels from the target.



The problem of conventional tracking algorithms

SAD template matching

-> ->




Conventional Mean Shift

->



When objects contain apertures, background pixels will be mixed into the target and degrade the purity of the target features (hereafter, we call it as the background interfusion problem).

Conventional tracking algorithms such as template matching and mean-shift algorithm suffer from this problem.



Our solution to the background interfusion:

We apply K-means clustering algorithm into the selected target and background samples in a 5D space to discriminate the interfused background from the target object.
We use the following ideas to achieve the robust object tracking:

(1) Representing the image feature with a constructed 5D feature vector.






(2) Applying K-means clustering to the target and background samples simultaneously.



We select the background samples from the ellipse contour automatically and apply K-means clustering between the target and background information to classify if an unknown pixel is a target one or not.




(3) Continuously updating the search area according to the distribution of the target pixels.





Experimental result of tracking non-rigid object.
   




(4) Self-tracking failure detection and recovery




Previous image Tracking failure occurs and the ellipse contour turns green to indicate it.

By searching over the ellipse and find the nearest points to the previous target centers in the 5D space, the tracking failure can be recovered.



Experiments

Comparative experiment with the rigid object containing apertures

For Mean-shift method and Template Matching method, tracking failed under the influence of the background through apertures.





Comparative experiment with non-rigid object containing apertures


The non rigid deformation of target causes the background pixels to be mixed into the search area. Such pixels will change the shape of color histogram in mean-shift algorithm. As for the template matching method, these pixels also change the grey value of the search area.





Some experimental results of our K-means tracker with various target objects
(including rigid and non-rigid objects)



-> ->
The area of target colors begin to change.
The appearance of target object almost completely changes.




-> ->
Motion blur and the reflection of glasses occurred at the same time.
Blinking.




-> ->
The target person began to disguise. The target person has completely disguised.





-> ->





-> ->





-> ->





-> ->





-> ->








Application of the K-means tracker

Combining K-means tracker with active cameras to capture clear images.


Demonstration of applying K-means tracker to control the active camera system,
on ICCV2005 Beijing, October 20th, 2005







連絡先:

〒640-8510 和歌山市栄谷930番地 和歌山大学 システム工学部 A603 
TEL&FAX: 073-457-8096

copyright (c) 2007 VRL - Vision & Robotics Laboratory - All Rights Reserved.