varun-projects / radar-to-lidar-place-recognition

Radar-to-Lidar: Heterogeneous Place Recognition via Joint Learning

radar-to-lidar-place-recognition

This page is the implementation of a pre-print, implemented by PyTorch.

Method

Data

The files in matlab/RobotCar_data and matlab/MulRan_data can help you generate scancontext of radar and lidar submaps. Also, the generation of lidar submaps is included.

Training

The train_disco_lidar_quad.py is used for training lidar-to-lidar DiSCO.

The train_disco_radar_quad.py is used for training radar-to-radar DiSCO.

The train_joint_radar_lidar.py is used for training L2L, R2R and R2L jointly based on DiSCO implementation.

The trained models are listed in the trained_models respectively.

Inference

Please use the files in inference folder.

Evaluation

In addition, the matlab/evaluate_recall@1 contains the files to calculate the recall@1 for place recognition evaluation.

Case Example

Multi-session place recognition: radar-to-lidar in different days of Mulran-Riverside

Citation

If you use our code in an academic work or inspired by our method, please consider citing the following:

@article{yin2021radar,
  title={Radar-to-Lidar: Heterogeneous Place Recognition via Joint Learning},
  author={Yin, Huan and Xu, Xuecheng and Wang, Yue and Xiong, Rong},
  journal={Frontiers in Robotics and AI},
  year={2021},
  status={Accept}
  }

And also, another related implementation is avaliable at DiSCO.

We also propose an end-to-end radar tracking method on lidar maps. Please refer to RaLL for details.

TODO

Make the original data and lidar filter files avaliable.