FragTrack: Fragments-based Tracking

FragTrack is part of the Visual Object Tracking Repository, which aims at providing a central repository for state-of-the-art tracking algorithms that are freely available. The source code for this tracker was obtained from its project website and extended by a challenge mode. The following description was copied literally from the original author.

README

By: Amit Adam

amita@cs.technion.ac.il
www.cs.technion.ac.il/~amita

Date: November 18th, 2007

---

General

This distribution contains the source code for a fragments-based tracker. It is written in C++ and uses the OpenCV library.

What's in the package

1. Fragments_Tracker.h,cpp - the tracker object code
2. fragtrack_envelope.cpp - an envelope for running the tracker on an image sequence
3. emd.h,cpp - code for comparing two histograms using Earth Mover's Distance - courtesy of Yossi Rubner
4. A Visual Studio solution for building the project.
5. Sample setup files for two image sequence. The sequences may be found in my homepage.
6. Sample log file.

Usage

1. Build the executable - a console application (tested only in "release configuration")

2. Prepare a setup file called "setup.txt" and place it in the same directory as the executable file.

3. Run the executable.

4. The following output should be obtained:

   • during the run - an OpenCV window with the tracking results
   • a log file called "FragTrack_log.txt" containing the tracking results
   • two images "initial_temlate.jpg" and "initial_target.jpg" showing the initial template

Format of setup file

The setup file is a text file containing 7 lines in the following format:

F:\amita\data\face_sequence\ % line 1 - path and file name prefix 1 % line 2 - first image in sequence 890 % line 3 - last image in sequence 75 120 220 235 % line 4 - target position in first image 7 % line 5 - search window half size 16 % line 6 - number of bins in histogram 3 % line 7 - choice of metric for comparing histograms

(do not include the comments in the setup file)

Here are some details:

The first 3 lines specify where to find the input sequence - in the above example the sequence is F:\amita\data\face_sequence\ 1.jpg, F:\amita\data\face_sequence\2.jpg, ..., F:\amita\data\face_sequence\890.jpg Note: no spaces are allowed in path or file name

Line number 4 gives the top-left and bottom-right corners of the target position in the first frame. The y-coordinate (row) is given first: tl_y tl_x br_y br_x

Line number 5 specifies the search radius around the position in previous frame (in pixels).

The algorithm is based on gray-scale intensity histograms. Line number 6 specifies the number of bins in the histograms.

For comparing two histograms the algorithm currently uses one of three options. Line 7 specifies which option: 1 means chi-square metric, 2 means EMD metric, 3 means a variation of the Kolmogorov-Smirnov statistic. The EMD is a cross-bin metric in contrast with standard bin-to-bin metrics such as Chi square. For one dimensional data option 3 is a much faster equivalent to the EMD metric. Option 3 should be your default choice. You can see the advantage option 3 has over option 1 on the "woman" sequence for example.

Example setup files and sequences

Two example setup files are contained in the distribution. The corresponding sequences are available from

www.cs.technion.ac.il/~amita

together with a file containing the ground truth for these sequences.

Feedback

Feedback (both positive and negative) is most welcome. Please email to amita@cs.technion.ac.il

Acknowledgement

Thanks to Yossi Rubner for his EMD code and for permission to redistribute it with this package.

Reference

Amit Adam, Ehud Rivlin, Ilan Shimshoni: Robust Fragments-based Tracking using the Integral Histogram. Proc. CVPR 2006, pp. 798-805