HKUST-Aerial-Robotics/MonoLaneMapping

Introduction

This is the official code repository of "Online Monocular Lane Mapping Using Catmull-Rom Spline", which is accepted by IROS'23.

MonoLaM (Monocular Lane Mapping) is an online lane mapping algorithm based on a monocular camera. It takes real-time images and odometry (such as from VIO), and estimates its own pose as well as the lane map.

Its main features are:

MonoLaM uses a monocular 3D lane detection network to obtain 3D lane marking measurements.
Lane association using a combination of Chamfer distance, pose uncertainty, and lateral sequence consistency.
The lane marking is parameterized by the Catmull-Rom Spline, which saves the map memory.
Initializing control points for unordered lane line point clouds.
Incrementally extending and optimizing the lane map.

Prerequisites

ROS

Follow the official guide to install ROS1.

Python Dependencies

pip install -r requirements.txt

GTSAM

Follow the official guide to install GTSAM with Python Bindings.

OpenLane ROS Package

mkdir -p catkin_ws/src
cd catkin_ws/src
git clone https://github.com/qiaozhijian/openlane_bag.git

Build

git clone https://github.com/HKUST-Aerial-Robotics/MonoLaneMapping.git
cd ..
catkin_make

OpenLane Example

We validate the method based on the OpenLane dataset. Users need to download this dataset and the rosbags we provide after preprocessing with openlane_bag.

Specifically, we use PersFormer to predict 3D lane markings and save them along with GT and calibration parameters (Details). The camera coord sys (OpenLane) is x-front, y-left, z-up. However, you may feel confused if you want preprocess by yourself using PersFormer. Because its camera coord sys is different from OpenLane (please refer to issue 24.

Rosbag download link [OneDrive][Baidu Cloud]

Unzip the downloaded file and put it in the OpenLane dataset folder.

├── OpenLane
│   └── lane3d_1000
│       ├── rosbag
│       └── test
│       └── validation
│       └── training

Modify the config/lane_mapping.yaml file to change the dataset path.

dataset:
    name: "openlane"
    dataset_dir: "/media/qzj/Document/datasets/OpenLane/"

Quick Start

python examples/demo_mapping.py --cfg_file=config/lane_mapping.yaml

Reproduce the results in the paper

#In this step, users need to download the original [OpenLane](https://github.com/OpenDriveLab/OpenLane) dataset and merge it with our provided dataset.
cd src/MonoLaneMapping
# lane mapping and save the results
python examples/mapping_bm.py --cfg_file=config/lane_mapping.yaml
# evaluation of lane recall and precision
python examples/openlane_eval3d.py --cfg_file=config/lane_mapping.yaml
# lane association evaluation
python examples/lane_association.py --cfg_file config/lane_association.yaml --bm

Toy example for curve fitting

python examples/demo_curve_fitting.py

Citation

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

@inproceedings{qiao2023online,
  title={Online monocular lane mapping using catmull-rom spline},
  author={Qiao, Zhijian and Yu, Zehuan and Yin, Huan and Shen, Shaojie},
  booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={7179--7186},
  year={2023},
  organization={IEEE}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

HKUST-Aerial-Robotics / MonoLaneMapping