V2I-CALIB and V2I-CALIB++: Object-Level, Real-Time Point Cloud Global Registration Framework for V2I/V2X Applications

V2X-Calib Paper | V2X-Calib++ Paper

Highlight

• An initial-value-free online calibration method for vehicle-road multi-end scenarios is proposed, based on perception objects;
• A new multi-end target association method is proposed, which fully explores spatial associations in the scene without positioning priors;
• oIoU and oDist both enable real-time monitoring of external parameters in the scene.

News

• [2024/09/13] V2I-CALIB++ is available here.
• [2024/06/30] V2I-CALIB is accepted by IROS 2024!

Experimental Comparison

We conducted experiments comparing V2I-Calib and V2I-Calib++ against well-performed point cloud Global Registration methods, using two widely recognized V2X datasets: DAIR-V2X and V2X-Sim. The results are as follows.

• V2X-Sim (Synchronous Homologous LiDARs)

  Method	RRE(°)	RTE(m)	Success Rate(%)		Time (s)
  			@1	@2
  FGR	0.69	0.16	78.64	95.15	0.92
  Quatro	0.17	0.18	96.40	98.20	0.83
  Teaser++	0.77	0.17	76.70	94.17	0.91
  V2I-Calib(Ours)	0.06	0.03	93.26	95.48	0.37
  V2I-Calib++(Ours)	0.01	0.01	96.80	98.31	0.13

  * Note: @λ indicates the threshold for success rate

• DAIR-V2X (Asynchronous Heterogeneous LiDARs)

  Method	RRE(°)	RTE(m)	Success Rate(%)		Time (s)
  			@1	@2
  FGR	1.71	1.61	22.11	62.81	25.50
  Quatro	1.46	1.49	19.90	69.90	24.52
  Teaser++	1.83	1.67	20.30	58.91	24.33
  V2I-Calib(Ours)	1.25	1.32	42.81	76.04	0.34
  V2I-Calib++(Ours)	1.23	1.16	51.40	84.58	0.12

  * Note: @λ indicates the threshold for success rate

Getting Started

Installation

This code is mainly developed under Ubuntu 20.04. We use anaconda3 with Python 3.8 as the base Python setup.

After cloning this repo, please run:

source setup.sh

Minimal Test

To test the sample, simply run the following command:

python test.py --test_type single 

Batch Test

For batch testing, additional data preparation is required. This process is also included in the test.py file.

Download data and organize as follows

Download DAIR-V2X-C dataset here and organize as follows:

# For DAIR-V2X-C Dataset located at ${DAIR-V2X-C_DATASET_ROOT}
├── cooperative-vehicle-infrastructure      # DAIR-V2X-C
    ├── infrastructure-side             # DAIR-V2X-C-I
        ├── velodyne                
            ├── {id}.pcd   
        ├── label   
            ├── camera                  # Labeled data in Infrastructure Virtual LiDAR Coordinate System fitting objects in image based on image frame time
                ├── {id}.json
            ├── virtuallidar            # Labeled data in Infrastructure Virtual LiDAR Coordinate System fitting objects in point cloud based on point cloud frame time
                ├── {id}.json
        ├── data_info.json              # Relevant index information of Infrastructure data
    ├── vehicle-side                    # DAIR-V2X-C-V
        ├── velodyne             
            ├── {id}.pcd
        ├── label   
            ├── camera                  # Labeled data in Vehicle LiDAR Coordinate System fitting objects in image based on image frame time
                ├── {id}.json
            ├── lidar                   # Labeled data in Vehicle LiDAR Coordinate System fitting objects in point cloud based on point cloud frame time
                ├── {id}.json
        ├── data_info.json              # Relevant index information of the Vehicle data
    ├── cooperative                     # Coopetative Files
        ├── label_world                 # Vehicle-Infrastructure Cooperative (VIC) Annotation files
            ├── {id}.json           
        ├── calib
            ├── lidar_i2v               # External Parameters from Infrastructure LiDAR to Vehicle LiDAR
                ├── {id}.json           # Vehicle ID
        ├── data_info.json              # Relevant index information combined the Infrastructure data and the Vehicle data

Note: cooperative-vehicle-infrastructure/cooperative/calib/lidar_i2v is generated by https://github.com/AIR-THU/DAIR-V2X/blob/main/tools/dataset_converter/calib_i2v.py.

Create a symlink to the dataset root

cd ${v2i-calib_root}/v2i-calib
mkdir ./data/DAIR-V2X
ln -s ${DAIR-V2X-C_DATASET_ROOT}/cooperative-vehicle-infrastructure ${v2i-calib_root}/v2i-calib/data/DAIR-V2X

Run Test Command

python test.py --test_type batch

The results are detailed in Log/xx.log. Execute Log/analyze.py to analyze the batch test results. The final results will be available in analysis_results.csv. You will find these results to be superior to those previously discussed or those presented in the paper :-)

Explanation of Key Configuration Parameters

The configuration parameters are located in config/config.yaml. To use the oIoU metric, set core_similarity_component_list = [iou]. To use the oDist metric, set core_similarity_component_list = [centerpoint_distance, vertex_distance]

Acknowledgment

This project is not possible without the following codebases.

• DAIR-V2X
• LiDAR-Registration-Benchmark

Citation

If you find our work or this repo useful, please cite:

@article{qu2024v2i,
  title={V2I-Calib: A Novel Calibration Approach for Collaborative Vehicle and Infrastructure LiDAR Systems},
  author={Qu, Qianxin and Xiong, Yijin and Wu, Xin and Li, Hanyu and Guo, Shichun},
  journal={arXiv preprint arXiv:2407.10195},
  year={2024}
}

@article{qu2024v2iplus,
  title={V2I-Calib++: A Multi-terminal Spatial Calibration Approach in Urban Intersections for Collaborative Perception},
  author={Qu, Qianxin and Zhang, Xinyu and Xiong, Yijin and Guo, Shichun and Song, Ziqiang and Li, Jun},
  journal={arXiv preprint arXiv:2410.11008},
  year={2024}
}