RobMOT: Robust 3D Multi-Object Tracking Framework

Official implementation of "RobMOT: Robust 3D Multi-Object Tracking by Observational Noise and State Estimation Drift Mitigation on LiDAR PointCloud" (Online Method).

Key Features

🚀 State Estimation Refinement: Novel Kalman Filter adaptation reduces localization noise by 4.8% HOTA
👻 Ghost Track Mitigation: Introduces the first trajectory validity mechanism in object tracking, reduces false positives by 80%
⚡ Real-Time Performance: 3221 FPS on single CPU - 50x faster than SOTA methods
🔄 Occlusion Handling: Improved recovery after prolonged occlusions with uncertainty-aware tracking
📊 Multi-Detector Support

Benchmarks

Dataset	MOTA ↑	HOTA ↑	IDSW ↓
KITTI Test	91.02%	81.76%	7
KITTI Val	86.31%	91.53%	1

Waymo Test (Vehicle)	MOTA/L1 ↑	MOTA/L2 ↑
All	77.72%	74.66%
Range-[0.30)	92.17%	91.59%
Range-[30, 50)	78.78%	75.99%
Range-[50,+inf)	59.74%	55.14%

Installation

Requirements

OS: Ubuntu 20.04/22.04 (64-bit)
Compiler: GCC 13+ (C++17 required)
CMake: 3.28+
Dependencies:
- OpenCV 4.1+
- PCL 1.14+
- Eigen3 3.4+

# Install dependencies
sudo apt-get install -y \
    gcc-13 g++-13 \
    libopencv-dev \
    libpcl-dev \
    libeigen3-dev

# Clone repository
git clone https://github.com/yourusername/RobMOT.git
cd RobMOT

Data Preparation

Organize the dataset as: (**Note the code does not include visualization, you may exclude directories with "# No need with no visualization")

datasets/
├── kitti/           
    ├── Camera # No need with no visualization
        ├── training 
            ├── 0000
            ...
        ├── testing
            ├── 0000
            ...
    ├── LiDAR
        ├── training
            ├── calib
                ├── 0000.txt
                ...
            ├── pose
                ├── 0000.txt
                ...
            ├── velodyne # No need with no visualization
                ├── 0000
                ...
        ├── testing
            ├── calib
                ├── 0000.txt
                ...
            ├── pose
                ├── 0000.txt
                ... 
            ├── velodyne # No need with no visualization
                ├── 0000
                ..

├── waymo/
    ├── Camera # No need with no visualization
        ├── training
            ├── 17065833287841703_2980_000_3000_000
            ...
        ├── testing
            ├── 2601205676330128831_4880_000_4900_000.parquet
            ...
    ├── LiDAR
        ├── training
            ├── pose
                ├── 17065833287841703_2980_000_3000_000.txt
                ...
            ├── velodyne # No need with no visualization
                ├── 17065833287841703_2980_000_3000_000
                ...
        ├── testing
            ├── pose
                ├── 17065833287841703_2980_000_3000_000.txt
                ...
            ├── velodyne # No need with no visualization
                ├── 17065833287841703_2980_000_3000_000
                ...

Configuration

All directories should be set up correctly to avoid issues when running the software. Please note all detector data are in folder advanced_detection.

config/config.cpp

        //TODO: Include the fill project path. The path should end by 'RobMOT/src'
        constexpr static const char* basePath  = "/EXAMPLE/RobMOT/src"; # The directory where you placed the repository, update "/EXAMPLE" only.

main.cpp

            const string BASE_DIR               = "EXAMPLE/datasets/"; // TODO: Update to the dataset directory on your machine, update "/EXAMPLE" only.

Configuration Options

config/config.cpp

        # select dataset -> values: kitti or waymo
        constexpr static const char* dataset_name = "kitti";
        # select dataset type -> values: training or testing
        constexpr static const char* dataset = "training";
        # true/false based on the dataset type.   
        constexpr static const bool  isTraining                 = true;
        # To not include state estimation in the results when objects are occluded. (Note: some datasets do not have ground truth for occluded objects)
        constexpr static const bool  processDetectedObjectsOnly = true;
        # To run a specific stream. Add a single or multiple stream names to the list, and put true for the `selectedStreamsActivation`
        constexpr static const bool selectedStreamsActivation   = false; // To run specific stream/s
        constexpr static const char* selectedStreams[1]   ={"0002"};

....
        // Define selectedDetector here
        # Change the detector name, Options:
        # for kitti (Virconv, Casc, Pointrcnn, Pvcnn, SecondIou),
        # for waymo (CascWaymo, Ctrl)
        inline std::unique_ptr<parameters::Detector> selectedDetector = std::make_unique<Virconv>();

Citation

@ARTICLE{11071990,
  author={Nagy, Mohamed and Werghi, Naoufel and Hassan, Bilal and Dias, Jorge and Khonji, Majid},
  journal={IEEE Transactions on Intelligent Transportation Systems}, 
  title={RobMOT: 3D Multi-Object Tracking Enhancement Through Observational Noise and State Estimation Drift Mitigation in LiDAR Point Clouds}, 
  year={2025},
  volume={},
  number={},
  pages={1-13},
  keywords={3D multi-object tracking;state estimation;Kalman filter;LiDAR point cloud},
  doi={10.1109/TITS.2025.3581980}}

MohamedNagyMostafa/RobMOT

README