Need help with mahalanobis_3d_multi_object_tracking?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

215 Stars 50 Forks Other 9 Commits 5 Opened issues

Services available


Need anything else?

Contributors list

No Data

Probabilistic 3D Multi-Object Tracking for Autonomous Driving

Hsu-kuang Chiu1, Antonio Prioletti2, Jie Li2, Jeannette Bohg1

1Stanford University, 2Toyota Research Institute

First Place Award, NuScenes Tracking Challenge, at AI Driving Olympics Workshop, NeurIPS 2019.


We present our on-line tracking method, which wins the first place award of the NuScenes Tracking Challenge[1], held in the AI Driving Olympics Workshop at NeurIPS 2019. Our technical report is available in arxiv. We also provide our source code and the step-by-step instructions to reproduce the validation set performance numbers.

Quantitative Results

Validation Set Results

The following table shows our quantitative tracking results for the validation set of NuScenes: evaluation in terms of overall Average Multi-Object Tracking Accuracy (AMOTA) and individual AMOTA for each object category in comparison with the tracking challenge official AB3DMOT[2] baseline results. The AB3DMOT[2] baseline and our method use the same MEGVII[3] detection results as the input of the tracking systems. For the full validation set results including our implementation of the AB3DMOT[2] baseline and our ablations, please refer to our paper in arxiv.


Overall bicycle bus car motorcycle pedestrian trailer truck
AB3DMOT[2] 17.9 0.9 48.9 36.0 5.1 9.1 11.1 14.2
Our proposed method 56.1 27.2 74.1 73.5 50.6 75.5 33.7 58.0

We can see that our method improves the AMOTAs significantly, especially for the smaller objects, such as pedestrians.

Test Set Results

The NuScenes Tracking Challenge organizer shared the test set performance of the top 3 participants and the official AB3DMOT[2] baseline, as shown in the following table. The AB3DMOT[2] baseline and our method use the same MEGVII[3] detection results as the input of the tracking systems. The full tracking challenge leaderboard can be seen at the tracking challenge webpage.


Team Name Overall
1 StanfordIPRL-TRI (Ours) 55.0
2 VV_team 37.1
3 CenterTrack 10.8
baseline AB3DMOT[2] 15.1

Qualitative Results

The following figures are the bird-eye-view visualization of the tracking results from the official AB3DMOT[2] baseline and our method. For this scene, we draw all the car bounding boxes from different timesteps of the same scene in a single plot. Different colors represent different tracking ids. We also show the ground-truth annotations and the input detections as the references. For more visualization results, please refer to our paper in arxiv.



Input Detections


We can see that our method is able to better track the object making a sharp turn.

Using Our Code to Reproduce the Validation Set Results

  1. Choose a directory YOURWORKINGDIR (mine is /juno/u/hkchiu), and download our code and setup the environment. In the following command, replace /juno/u/hkchiu with YOURWORKINGDIR:

    cd /juno/u/hkchiu
    git clone 
    conda create --name probabilistic_tracking python=3.6
    conda activate probabilistic_tracking
    cd mahalanobis_3d_multi_object_tracking
    pip install -r requirements.txt
  2. Choose a directory YOURNUSCENESDEVKITDIR (mine is /juno/u/hkchiu) and download the nuscenes devkit. In the following command, replace /juno/u/hkchiu with YOURNUSCENESDEVKITDIR:

    cd /juno/u/hkchiu
    git clone
    cd nuscenes-devkit
    pip install -r setup/requirements.txt
  3. Include nuscenes devkit into python path. In the following command, replace /juno/u/hkchiu with YOURNUSCENESDEVKIT_DIR. Note that you will need to run this everytime after you activate the conda environment):

    export PYTHONPATH="${PYTHONPATH}:/juno/u/hkchiu/nuscenes-devkit/python-sdk"
  4. Choose a directory YOURNUSCENESDATA_DIR (mine is /juno/u/hkchiu/dataset/nuscenes), and download then unzip the mini, trainval, and test data from the NuScenes download webpage. Note that you will need to create an account there. The total size of the zip files is roughly 360GB, and the unzip step also requires the same amount of extra disk space.

  5. Choose a directory YOURNUSCENESDETECTIONDIR (mine is /juno/u/hkchiu/dataset/nuscenesnew), and download then unzip the MEGVII[3]'s detection results from the NuScenes tracking webpage.

  6. Edit our code: in,,, replace “/juno/u/hkchiu/dataset/nuscenesnew/” with YOURNUSCENESDETECTIONDIR, and replace “/juno/u/hkchiu/dataset/nuscenes/” with your YOURNUSCENESDATADIR.

  7. Run the tracking code of our proposed method and the evaluation code (tracking takes 15 minutes, evaluation takes 90 minutes):

    python val 2 m 11 greedy true nuscenes results/000008;
    python --output_dir results/000008 results/000008/val/results_val_probabilistic_tracking.json > results/000008/output.txt
  8. You can also run our implementation of the AB3DMOT[2] baseline method:

    python val 0 iou 0.1 h false nuscenes results/000001;
    python --output_dir results/000001 results/000001/val/results_val_probabilistic_tracking.json > results/000001/output.txt
  9. Or you can run all the tracking methods (our implementation of AB3DMOT[2], our ablation methods, and our final proposed method) using the script:

  10. (Optional) You can also run to see how we use the training set data to estimate the Kalman Filter's covariance matrices. (10 minutes) I have hardcoded the results into the file, so running is not required to run the above tracking and evaluation commands.

    python > nuscenes_train_stats.txt


  • We implemented our method on top of AB3DMOT[2]'s open-source code.

  • We used MEGVII[3]'s detection results as the input of our tracking method.

  • Toyota Research Institute ("TRI") provided funds to assist the authors with their research but this article solely reflects the opinions and conclusions of its authors and not TRI or any other Toyota entity.


  • [1] "nuScenes: A multimodal dataset for autonomous driving", Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom, arXiv:1903.11027, 2019.
  • [2] "A Baseline for 3D Multi-Object Tracking", Xinshuo Weng and Kris Kitani, arXiv:1907.03961, 2019.

  • [3] "Class-balanced Grouping and Sampling for Point Cloud 3D Object Detection", Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, Gang Yu, arXiv:1908.09492, 2019.

To Cite Our Paper

  title={Probabilistic 3D Multi-Object Tracking for Autonomous Driving},
  author={Chiu, Hsu-kuang and Prioletti, Antonio and Li, Jie and Bohg, Jeannette},
  journal={arXiv preprint arXiv:2001.05673},

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.