RadarHD creates high resolution lidar-like point clouds from just a single-chip, cheap mmWave radar. This enables high quality perception even in vision/lidar denied scenarios such as in smoke or fog. For example: think of a futuristic fire fighting robot performing search and rescue in a smoky environment. This repository hosts files pertaining to this project that appeared in IEEE ICRA 2023.

Paper link | Demo link | Project website

RadarHD Overview

Pre-requisites

Install Docker.
Clone this repository at project_root.
Download pre-trained model. Move this to logs/13_1_20220320-034822/ folder in the cloned repository.
[Optional] Download the raw radar and lidar full-dataset (part1-dataset, part2-dataset, part3-dataset, part4-dataset) captured along 44 different trajectories. You can visualize each trajectory and map here. dataset_5 contains a processed version of this raw dataset to help users train and test quickly without needing to deal with the raw dataset. If you find the dataset too big to download, here is a smaller dataset.
Matlab (Only for point cloud error evaluation).

Repository structure

install.sh installs all dependencies.
train_radarhd.py and test_radarhd.py are used for training and testing our models.
Pre-trained model is stored in logs/13_1_20220320-034822/. This model was trained using radar-lidar images dataset in dataset_5/.
train_test_utils/ contains model, loss and dataloading definitions for training and testing.
eval/ contains scripts for evaluating RadarHD's generated upsampled radar images.
create_dataset/ contains scripts that show our pre-ML radar and lidar processing on raw sensor data. Use this only for creating your own radar-lidar images dataset (similar to dataset_5) to train with our models.

Usage

Create a Docker environment

  sudo docker run -it --rm --gpus all --shm-size 8G -v project_root:/radarhd/ pytorch/pytorch bash

Install all dependencies
```
  cd /radarhd/
  sh install.sh
```
- When prompted, be sure to setup time zone as US Eastern. This avoids any time zone related issues that show up during dataset creation.
For testing on pre-trained model logs/13_1_20220320-034822/ and test images in dataset_5/test/
```
  python3 test_radarhd.py
```
- For testing with other models and datasets, modify the constants in test_radarhd.py.
- To test on CPU, make sure to use CPU device.
For training using params similar to logs/13_1_20220320-034822/ and train images in dataset_5/train/
```
  python3 train_radarhd.py
```
- For training with your own params and datasets, modify the constants in train_radarhd.py
For evaluating the output of test_radarhd.py:
- Executing test_radahd.py will create generated upsampled radar and ground truth lidar images in polar format for all the test data in the corresponding log folder. (Default: logs/13_1_20220320-034822/test_imgs/)
- Convert polar images to cartesian.
```
  cd ./eval/
  python3 pol_to_cart.py
```
- Convert cartesian images to point cloud for point cloud error evaluation.
```
  python3 image_to_pcd.py
```
- Visualize the generated point clouds for qualitative comparison in Matlab.
```
 pc_vizualize.m
```
- Generate quantitative point cloud comparison in Matlab (similar to eval/cdf.jpg)
```
  pc_compare.m
```

Citation

If you found this work useful, please consider citing this work as:

@INPROCEEDINGS{10161429,
author={Prabhakara, Akarsh and Jin, Tao and Das, Arnav and Bhatt, Gantavya and Kumari, Lilly and Soltanaghai, Elahe and Bilmes, Jeff and Kumar, Swarun and Rowe, Anthony},
booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)}, 
title={High Resolution Point Clouds from mmWave Radar}, 
year={2023},
volume={},
number={},
pages={4135-4142},
doi={10.1109/ICRA48891.2023.10161429}}

akarsh-prabhakara/RadarHD

RadarHD Overview

Pre-requisites

Repository structure

Usage

Citation