📌 This paper has been accepted for publication in IEEE Robotics and Automation Letters (RA-L), 2025.

Accurate Pose Refinement of Detected Vehicles Using LiDAR Point-to-Surfel ICP and Vehicle Shape Priors

Repository for paper "Accurate Pose Refinement of Detected Vehicles Using LiDAR Point-to-Surfel ICP and Vehicle Shape Priors"

This repository provides code of prior model-based detection refinement moudle, dataset and evaluation tool. (Our codes will be provided soon. Stay tuned!)

System architecture

Detection improvement (mIoU) results by various voxel size

Method

1. Downsample vehicle prior model into Surfel Model

Downsampling CAD model to Surfel model

2. Execute 4-DoF Point-to-Surfel registration.

Refinement process

Demo

• Stop Scenario Comparision

(Left) Tracking without refinement : Due to unstable detection, static objects become dynamic (Right) Tracking with refinement : Static due to stable detection

Run Algorithm

1. Detection only

roslaunch box_point_generator box_point_generator.launch
roslaunch cad_registration cad_registration.launch

2. Run Tracking (TBD)

Evaluation

Evaluation Setup & Sensors

• LiDAR : Velodyne VLP-32C
• GNSS : Novtel CPT-7, Novatel Pwrpak7

Evaluation setup and test track

Scenario Example

4 scenarios

1. Stop Scenario (18.3 sec): Target vehicles static. Ego vehicle moves slowly.
2. Slow Scenario (30.0 sec): The target vehicles and the ego vehicle move slowly in parallel.
3. Fast Scenario short (159 sec) : Target vehicles and ego vehicles drive at high speed, overtaking each other
4. Fast Scenario long (399 sec) : Target vehicles and ego vehicles drive at high speed, overtaking each other

Scenario Bag file

[Google Drive Link] will be provided soon. Stay tuned!

1. Stop Scenario
   1. Ego: /Ego_Vehicle/stop_3_ego.bag
   2. Target 1: /Target_Vehicle_1/target_1_3.bag
   3. Target 2: /Target_Vehicle_2/target_2_3.bag
2. Slow Scenario
   1. Ego: /Ego_Vehicle/slow_0_ego.bag
   2. Target 1: /Target_Vehicle_1/target_1_0.bag
   3. Target 2: /Target_Vehicle_2/target_2_0.bag
3. Fast Scenario short
   1. Ego: /Ego_Vehicle/fast_3_short_ego.bag
   2. Target 1: /Target_Vehicle_2/target_1_3.bag
   3. Target 2: /Target_Vehicle_2/target_2_3.bag
4. Fast Scenario long
   1. Ego: /Ego_Vehicle/fast_2_long_ego.bag
   2. Target 1: /Target_Vehicle_2/target_1_2.bag
   3. Target 2: /Target_Vehicle_2/target_2_2.bag

Data Preparation

1. Need GNSS Topic from target vehicle (Existed in bag files)

• /novatel/oem7/inspvax

2. Ego detection topoic type should be "autohyu_msgs::DetectsObjects" (Will be change to a general topic type.)

• topic_name/ego_detection_topic_name

Run Evaluation Tool

0. Setup system.yaml file
1. Launch Evaluation tool

roslaunch detection_evaluation detection_evaluation.launch

2. rviz file in /rviz/autoku_evaluation.rviz

3. Type scene number when "Enter index of scene to visualize ('q' to quit)" pop up

Citation

If you find this work useful in your research, please cite:

@article{kim2025accurate,
  title={Accurate Pose Refinement of Detected Vehicles Using LiDAR Point-to-Surfel ICP and Vehicle Shape Priors},
  author={Kim, Soyeong and Jo, Jaeyoung and Lee, Jaehwan and Jo, Kichun},
  journal={IEEE Robotics and Automation Letters},
  year={2025},
  publisher={IEEE}
}

Contact

If you have any questions, please let me know:

• Soyeong Kim ([email protected]), Jaeyoung Jo ([email protected]), Jaehwan Lee
• AILAB Hanyang University (https://autolab.hanyang.ac.kr/)