An Online Multi-Object Tracking (MOT) Method with the GMPHD Filter and Occlusion Group Management (OGM).

To implement a highly practical MOT method, online approach and real-time speed are required.

• Online Process: The Gaussian mixture probability density (GMPHD) filter was exploited as a base model.
• Real-time Speed: We only used bounding box information (x, y, width, height).

Tracking-by-detection paradigm.

• In this project, the public detection results on the MOT15 (ACF) and MOT17 (DPM, FRCNN, SDP) datasets are used as inputs of the GMPHD-OGM tracker.

Our main contributions are as follows:

• An effective extension of the conventional GMMPHD filter from the original domain radar/sonar systems to video systems, consisting of:
  • A hierarchical data association strategy: detection-to-track and track-to-track associations.
  • An occlusion group management scheme: merge and occlusion energy minimization.
• The GMPHD-OGM tracker achieved the competitive tracking accurary and real-time speed (>30 FPS).
  • Experimental results was presented in [MOT15], [MOT17], and arxiv in detail.

Windows 10 (64 bit)
Visual Studio 2015 (64 bit)

Visual C/C++

OpenCV 3.4.1 and boost 1.61.0 (Windows Binaries) were used to implement the GMPHD-OGM tracker.

Download Win Pack and boost_1_61_0-msvc-14.0-64.exe to run our tracker in Visual Studio 2015 (64 bit).

PROJECT_HOME
├── GMPHD-OGM.sln  <-- **solution file for the project**
└── GMPHD-OGM      
    ├── demo_GMPHD-OGM.cpp                            <-- **the main function**
    ├── GMPHD_OGM.h, GMPHD_OGM.cpp                    <-- *a class implementation of the GMPHD-OGM tracker*
    ├── HungarianAlgorithm.h, HungarianAlgorithm.cpp  <-- a class implementation of the Hungarian Algorithm 
    ├── stdafx.h                                      <-- precompiled header including essential header files
    ├── GMPHD-OGM.vcxproj, GMPHD-OGM.vcxproj.filters  <-- VS project file, VS project filter file
    ├── params                                        <-- text files containing scene parameters
    |   └── MOT15train_ACF.txt, MOT17train_DPM.txt, MOT17train_FRCNN.txt, MOT17train_SDP.txt
    ├── seqs                                          <-- text files containing dataset paths and sequences' names
    |   └── MOT15train_ACF.txt, MOT17train_DPM.txt, MOT17train_FRCNN.txt, MOT17train_SDP.txt
    └── res                                           <-- tracking results are saved
        ├── MOT15
        |   └── test, train 
        └── MOT17
            └── test, train 

HungarianAlgorithm.h, HungarianAlgorithm.cpp

is a class implementation of the Hungarian Algorithm refering to #mcximing/hungarian-algorithm-cpp

1. Open the solution file (GMPHD-OGM.sln)
2. Link and include OpenCV3.4.1 and boost1.61.0 libraries to the project.
3. Press Ctrl+F5 in Release mode (64 bit)

Download the MOTChallenge datasets.

Users should specify the dataset path and sequences' names in the sequence list file.

e.g., GMPHD-OGM\seqs\MOT17train_DPM.txt
F:\MOT\MOTChallenge2017\train\
MOT17-02-FRCNN
MOT17-04-FRCNN
MOT17-05-FRCNN
MOT17-09-FRCNN
MOT17-11-FRCNN
MOT17-13-FRCNN

More details of the parameters will be updated soon.

e.g., GMPHD-OGM\params\MOT17train_DPM.txt

// demo_GMPHD-OGM.cpp
int main()
{
string seqList = "seqs\MOT17train_DPM.txt";
string seqList = "params\MOT17train_DPM.txt";
...

#define VISUALIZATION_MAIN_ON 0
#define SKIP_FRAME_BY_FRAME 0

VISUALIZATION_MAIN_ON: 0(off), 1(on)

• You can see the visualization windows of detection and tracking.
  • At initial frame, press any key to start tracking process.

SKIP_FRAME_BY_FRAME: 0(off), 1(on)

• You can see the results, frame-by-frame. (by pressing any key).

#define VISUALIZATION_MAIN_ON 1

[MOT17-05-SDP] Click this gif image for playing a full demo video

[MOT17-10-SDP] Click this gif image for playing a full demo video

#define VISUALIZATION_MAIN_ON 0

The tracking results files are saved as
[project home]\GMPHD-OGM\res[benchmark name][train or test][sequence name].txt

e.g., [project home]\GMPHD-OGM\res\MOT15\train\Venice-2.txt

We are only providing the parameters of the GMPHD-OGM tracker on the training sequences.

[1] Ba-Ngu Vo and Wing-Kin Ma, "The Gaussian Mixture Probability Hypothesis Density Filter," IEEE Trans. Signal Process., vol. 54, no. 11, pp. 4091–4104, Oct. 2006. [paper]

[2] Young-min Song and Moong Jeon, "Online Multiple Object Tracking with the Hierarchically Adopted GM-PHD Filter Using Motion and Appearance," in Proc. IEEE Int. Conf. Consum. Electron.-Asia (ICCE-Asia), pp. 1--4, Oct. 2016. [paper]

[3] Young-min Song, Young-chul Yoon, Kwangjin Yoon, and Moongu Jeon, "Online and Real-Time Tracking with the GM-PHD Filter using Group Management and Relative Motion Analysis," in Proc. IEEE Int. Workshop Traffic Street Surveill. Safety Secur. (AVSS), Nov. 2018. [paper]

[4] Young-mins Song, Kwangjin Yoon, Young-chul Yoon, Kin Choon Yow, and Moongu Jeon, "Online Multi-Object Tracking with GMPHD Filter and Occlusion Group Management," IEEE Access, vol. 7, pp. 165103-165121, Nov. 2019.

\bibitem{gmphdogm} Y. Song, K. Yoon, Y-C. Yoon, K. C. Yow, and M. Jeon, 
``Online Multi-Object Tracking with GMPHD Filter and Occlusion Group Management,'' 
IEEE Access, vol. 7, pp. 165103-165121, Nov. 2019.

BSD 2-Clause License