A Comparison of Mean Shift Tracking Methods

Authors

Nicole Maria Artner
Austrian Research Centers GmbH - ARC, Smart Systems Division, Vienna, Austria
nicole.artner@arc.ac.at

Wilhelm Burger
Digital Media, Upper Austria University of Applied Sciences, Hagenberg, Austria
wilhelm.burger@fh-hagenberg.at

Abstract

The mean shift algorithm is a well known statistical method to find local maxima in probability distributions. Besides filtering and segmentation it is applied in the field of object tracking. There are a lot of different approaches using mean shift to locate the target object in video frames. This paper explains three similar approaches and compares their performance with different test videos.

Keywords

object tracking, mean shift, CAMShift, weighted histogram, ratio histogram, candidate model

Paper

Download paper here.

Images and Videos of the experiments

The results are ordered as follows:

(1) CAMShift by Bradski
(2) Weighted histogram by Allan
(3) Ratio histogram by Allan
(4) Target and candidate model by Comaniciu

All videos are using the XVID codec. You can get it here http://www.xvid.org/.

Test video 1

(1)


Click here for video of result!

(2)


Click here for video of result!

(3)


Click here for video of result!

(4)


Click here for video of result!

Test video 2

(1)


Click here for video of result!

(2)


Click here for video of result!

(3)


Click here for video of result!

(4)


Click here for video of result!

Test video 3

(1)


Click here for video of result!

(2)


Click here for video of result!

(3)


Click here for video of result!

(4)


Click here for video of result!

Comparison to ground truth

The following graphs show the Euclidean distance of the results (position, width and height of object) of the four methods from ground truth. Ground truth was determined by hand for every tenth frame.

Test video 1


Test video 2


Test video 3