RMap: Millimter-Wave Radar Mapping Through Volumetric Upsampling

[IROS Paper] [Website] [Video]

RMap (Radar Mapping), a method to generate highprecision 3D maps using radar point clouds extracted from an mmWave sensor.

We present an end-to-end pipeline for generating the 3D maps from radar point clouds and demonstrate how these maps can be leveraged to construct a 3D map resembling lidar-based maps through UpPoinTr.

Usage:

1. For the coloRadar dataset, the maps generated using radar and lidar (also lidar_filtered - considering lidar measurements only in the range and FOV of radar). The maps are stored in data/ply. For the points along the trajectory, the data is stored in data/poses
2. From the maps and poses, generate radar input and lidar groundtruth patches by:

   python utils/poseSample.py --pcd_dir ./data/ply --input_dir <SAVE_INPUT_DIR> --gt_dir <SAVE_GT_DIR>
   

3. Train/Test the UpPoinTr network with the generated input (and gt) patches. More details are available in <a href="[./UpPoinTr](https://github.com/ajaymopidevi/UpPoinTr)">UpPoinTr</a> repo.
4. Combine the UpPoinTr predicted patches by

   python combinescenePCD.py
   

This saves the final combined map for scene and also outputs the CD-L1 and CD-L2 metrics

Pretrained Model and Data

1. Download AdaPoinTr pretrained weights: <a href = "https://drive.google.com/drive/folders/1ClTBiiSjXi-Xtuk7QJnS9p5R90vBnZoH?usp=sharing">GDrive</a>
2. Download ColoRadar data from <a href = "https://drive.google.com/drive/folders/15oWJZhAGHDBx9m_h7VacBsEcLbn288Ig?usp=sharing">GDrive</a> and upload in UpoinTr/data/ColoRadar/
3. Run inference

   cd UpoinTr
   python tools/inference.py cfgs/ColoRadar_models/AdaPoinTr.yaml ckpt_best.pth --pc_root data/ColoRadar/test --out_pc_root output/AdaPoinTr_FPSRadarLarge/ 
   

4. Generate combined maps

   python combineScenePCD.py
   

5. Generate scores to compare with lidar maps

   python generateScores.py
   

For generating radar maps on a new dataset:

1. Install <a href="https://github.com/1988kramer/octomap/tree/feature/intensity_map">octomap</a>
2. ROS package dependecies:
   • <a href="https://github.com/1988kramer/octomap_mapping/tree/feature/radar_image">ocotmap_mapping</a>
   • <a href="https://github.com/arpg/dca1000_device_msgs">dca1000_device_msgs</a>
   • <a href="https://github.com/Alphakyl/octomap_radar_analysis">octomap_radar_analysis</a>
3. Create a custom launch file similar to ocotomap_radar_analysis/launch/ocotmap_mapping.launch file

Results:

RMap genrated maps for ColoRadar dataset:

Through this crosssection analysis, we see that the original radar map consists primarily of noise. However, the RMap generated map has a similar structure to the lidar map, distinguishing between free space and occupied space.

Citation

If you find our work useful in your research, please consider citing:

@INPROCEEDINGS{10801827,
  author={Mopidevi, Ajay Narasimha and Harlow, Kyle and Heckman, Christoffer},
  booktitle={2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
  title={RMap: Millimeter-Wave Radar Mapping Through Volumetric Upsampling}, 
  year={2024},
  volume={},
  number={},
  pages={1108-1115},
  keywords={Laser radar;Three-dimensional displays;Simultaneous localization and mapping;Spaceborne radar;Radar;Millimeter wave radar;Transformers;Real-time systems;Trajectory;Odometry},
  doi={10.1109/IROS58592.2024.10801827}}