This repository provides the implementation of our Clean and Consistent MLS Point Clouds from the Start: Dynamic Removal Coupled with LIDAR-Inertial Odometry for Urban Scene, which improves pose estimation and mapping accuracy in dynamic urban environments. The system is model-driven, voxel-free, and supports seamless integration with state-of-the-art LIO frameworks.

✨ Our related papers has been submitted to "ISPRS Journal of Photogrammetry and Remote Sensing".

🚀 The source code will be released once our paper is accepted. Please stay tuned!

• Ubuntu: 20.04
• ROS: Noetic
  • Installation guide: ROS Installation

• PCL ≥ 1.8
  sudo apt install libpcl-dev

• Eigen ≥ 3.3.4

sudo apt install libeigen3-dev

Follow livox_ros_driver Installation.

Follow [TBB Installation] (Note: change the gcc-9.1/g++-9.1 to gcc-9/g++-9)

The City Datasets were utilized for the evaluation and ablation study of MA-LIO. This dataset was collected in urban environments using a car, exhibiting high velocity, dynamic objects, numerous rotations with U-turns, and tunnels. The dataset was acquired using three LiDAR sensors (Livox Avia, Livox Tele, and OS2-128) and a 100Hz IMU.

The KA-Urban Datasets were utilized for the evaluation and ablation study of LiLi-OM This dataset was collected in urban environments using a bicycle, exhibiting high vibrations and dynamic objects. The datasets recorded by Livox Horizon (10 Hz) and Xsens MTi-670 (200 Hz).

We built a MLS collection system, which integrates a Livox HAP solid-state LiDAR, a high-precision IMU, multiple monocular RGB cameras, and a tightly coupled GNSS/IMU navigation system for generating ground-truth poses. To achieve accurate temporal alignment among all sensors, precision time protocol was adopted, ensuring nanosecond-level synchronization between LiDAR, IMU, and GNSS data.

The four private datasets were collected in Shanghai, China, across four urban regions, each containing diverse dynamic activity patterns. The recorded routes (Private_01 to Private_04) span between several hundred meters and over one kilometer. These areas were intentionally selected to capture high scene dynamics, including moving cars, pedestrians, and bicyclists under natural urban conditions.

The source code is released under GPLv2 license.