PALoc: Advancing SLAM Benchmarking with Prior-Assisted 6-DoF Trajectory Generation and Uncertainty Estimation

Introduction

PALoc presents a novel approach for generating high-fidelity, dense 6-DoF ground truth (GT) trajectories, enhancing the evaluation of Simultaneous Localization and Mapping (SLAM) under diverse environmental conditions. This framework leverages prior maps to improve the accuracy of indoor and outdoor SLAM datasets. Key features include:

• Robustness in Degenerate Conditions: Exceptionally handles scenarios frequently encountered in SLAM datasets.
• Advanced Uncertainty Analysis: Detailed covariance derivation within factor graphs, enabling precise uncertainty propagation and pose analysis.
• Open-Source Toolbox: An open-source toolbox is provided for map evaluation, indirectly assessing trajectory precision.

News

• 2024/011/15: Docker support!
• 2024/08/15: Support newer college dataset!
• 2024/08/15: Support FusionPortable dataset and MS-dataset
• 2024/08/14: 🔥Release codes and data.
• 2024/03/26: Early access by IEEE/ASME TMECH.
• 2024/02/01: Preparing codes for release.
• 2024/01/29: Accepted by 2024 IEEE/ASME TMECH.
• 2023/12/08: Resubmitted.
• 2023/08/22: Reject and resubmitted.
• 2023/05/13: Submitted to IEEE/ASME TRANSACTIONS ON MECHATRONICS (TMECH).
• 2023/05/08: Accepted by ICRA 2023 Workshop on Future of Construction.

Dataset

GEODE Dataset

This data was provided by Zhiqiang Chen and Prof.Yuhua Qi from SYSU and HKU.

Stairs scenes with different types of lidar and glass noise. This is very challenging due to narrow stairs , you need to tune some parameters of ICP. The prior map and raw map can be downloaded.

Prior map without glass noise	Raw prior map

Sensor setup	Download link
Velodyne16+ xsense IMU	http://gofile.me/4jm56/yCBxjdEXA
Ouster64 + xsense IMU	http://gofile.me/4jm56/2EoKPDfKi

FusionPortable Dataset

Our algorithms were tested on the Fusion Portable Dataset.

Sequence	GT Map	Scene
20220216_corridor_day	corridor with x degeneracy.
20220216_canteen_day	The prior map only covers a portion of the scene.
20220219_MCR_normal_01	Performance on a quadruped robot platform.
20220216_escalator_day	Performance in an open stairwell scenario.
20220216_garden_day	Smaller scenario, similar to an indoor environment.
20220225_building_day	Three loops of indoor hallway scanning with a handheld device, taking a relatively long time.

Newer College Dataset

• Multicam Vision Lidar IMU dataset : Ouster 128 + Integrated IMU
• Stereo Vision Lidar IMU dataset: Ouster 64 + Integrated IMU

This dataset include 2 maps: parkland and math-institute.

Parkland	Math-institute

Self-collected Dataset

Below is our sensor kit setup.

Sequence	Scenes	GT
parkinglot_01		GT
redbird_02		GT

Getting Started

Install

• Open3d ( >= 0.11)
• PCL
• GTSAM 4.2.0 (fixed by @WangWenda98)

The GTSAM version should be compatible with running LIO-SAM smoothly. Additionally, there is no need to install the livox_ros_driver required by FAST-LIO2, as we have directly integrated the necessary message headers into the code.

Docker Support

It is recommended to run the code in the container while visualize it in the host machine, e.g.:

Pull the Docker image:

docker pull ulterzlw/paloc

Run the container with host network access:

docker run -it --network host ulterzlw/paloc bash

launch the application (inside the container):

roslaunch paloc geode_beta_os64.launch

Start RViz (on the host machine):

rviz -d ${PATH_TO_PALOC}/config/rviz/ouster_indoors.rviz

Quickly Run

download the demo rosbag and prior map, set the file path in geode_beta_os64.launch.

#ouster64 : GEODE dataset
roslaunch paloc geode_beta_os64.launch

#velodyne16: GEODE dataset
roslaunch paloc geode_alpha_vlp16.launch

#ouster128 fusionportable-corridor
roslaunch paloc fp_os128_corridor.launch

#pandar xt32: MS-dataset
roslaunch paloc ms_sbg_pandar32.launch

#ouster128: newer_colleage dataset
roslaunch paloc newer_colleage_os128.launch 

then play rosbag:

# ouster64
rosbag play stairs_bob.bag

#velodyne16
rosbag play stairs_alpha.bag

#ouster128 fusionportable-corridor
rosbag play 20220216_corridor_day_ref.bag

#pandar xt32
rosbag play Parkinglot-2023-10-28-18-59-01.bag 

#ouster128
rosbag play parkland0.bag

You can save data.

rosservice call /save_map

For your own dataset

Set Parameters

Set your file path in geode_beta_os64.launch. Put your prior map file in prior_map_directory, the map name must be set as the sequence. The map file size is recommend to down-sampled less than 2Gb.

<param name="save_directory" type="string" value="/home/xchu/data/paloc_result/"/>
<param name="prior_map_directory" type="string" value="/home/xchu/data/prior_map/paloc_map_file/"/>

<arg name="sequence" default="stairs_bob"/>

Adapt for FAST-LIO2

Set parameters in geode_beta_os64.yaml. Adapt for the FAST-LIO2 first.

common:
  lid_topic: "/ouster/points"
  imu_topic: "/imu/data"

  acc_cov: 1.1118983704388789e-01                     # acc noise and bias
  b_acc_cov: 1.5961182793700285e-03
  gyr_cov: 9.6134865171113148e-02                     # gyro noise and bias
  b_gyr_cov: 7.9993782046705285e-04

  extrinsic_T: [ -0.027172, -0.034873, 0.062643 ]     # from lidar to imu
  extrinsic_R: [ 0.998638, 0.052001, -0.004278,
                 -0.051937, 0.998554, 0.013900,
                 0.004994, -0.013659, 0.999894 ]
lio:
  lidar_type: 8      # 1 for Livox serials LiDAR, 2 for Velodyne LiDAR, 3 for ouster LiDAR,
  scan_line: 64
  scan_rate: 10      # only need to be set for velodyne, unit: Hz,
  timestamp_unit: 3  # 0-second, 1-milisecond, 2-microsecond, 3-nanosecond.
  blind: 0.5         # remove the nearest point cloud

Initial pose

then set the initial pose. When you run the launch command, the first point cloud will be saved in save_directory, you can align it with the prior map using CloudCompare to get the initial pose.

common:
    initial_pose: [ -0.519301, 0.850557, 0.082936 ,-11.347226,
                  -0.852164, -0.522691, 0.024698, 3.002144,
                  0.064357, -0.057849, 0.996249, -0.715776,
                  0.000000, 0.000000, 0.000000, 1.000000 ]

Evaluation with Cloud Map Evaluation

We can evaluate the map accuracy of PAloc as follows. Note that when you use the Cloud Map Evaluation library, the map of PALoc or ICP do not need to set initial pose since they are already in the same coordinate. But evaluate the map from FAST-LIO2 must to set it.

Raw Error Map	Entropy Map

Evaluation results example can be downloaded here.

TODO

• clean codes
• tutorial and parameters tuning
• support for LIO-SAM
• gravity factor
• adapt for more dataset and lidar

Results

Trajectory Evaluation

Map Evaluation

Degeneracy Analysis

X Degenerate (Translation)	Yaw Degenerate (Rotation)

Time Analysis

To plot the results, you can follow this scripts.

Important Issue

How do we calculate cov of odom factor in GTSAM?

We follow the assumption of pose Independence as barfoot does(TRO 2014) as equation (13), there is no correlated cov between this two poses. Left-hand convention means a small increments on world frame.

Since GTSAM follow the right-hand convention on SE(3) , we need to use the adjoint representation as equation (14). Please note that there is an error in Equation 14. The paper has not yet been updated. The adjoint matrix of the inverse relative pose should be used, not the adjoint of the relative pose.

Citations

For referencing our work in PALoc, please use:

@ARTICLE{hu2024paloc,
  author={Hu, Xiangcheng and Zheng, Linwei and Wu, Jin and Geng, Ruoyu and Yu, Yang and Wei, Hexiang and Tang, Xiaoyu and Wang, Lujia and Jiao, Jianhao and Liu, Ming},
  journal={IEEE/ASME Transactions on Mechatronics}, 
  title={PALoc: Advancing SLAM Benchmarking With Prior-Assisted 6-DoF Trajectory Generation and Uncertainty Estimation}, 
  year={2024},
  volume={},
  number={},
  pages={1-12},
  doi={10.1109/TMECH.2024.3362902}
  }

The map evaluation metrics of this work follow Cloud_Map_Evaluation. Please cite:

@article{jiao2024fp,
  author    = {Jianhao Jiao and Hexiang Wei and Tianshuai Hu and Xiangcheng Hu and Yilong Zhu and Zhijian He and Jin Wu and Jingwen Yu and Xupeng Xie and Huaiyang Huang and Ruoyu Geng and Lujia Wang and Ming Liu},
  title     = {FusionPortable: A Multi-Sensor Campus-Scene Dataset for Evaluation of Localization and Mapping Accuracy on Diverse Platforms},
  booktitle = {2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2022}
}

Acknowledgment

The code in this project is adapted from the following projects:

• The odometry method is adapted from FAST-LIO2 and LIO-SAM.
• The basic framework for pose graph optimization (PGO) is adapted from SC-A-LOAM.
• The Point-to-Plane registration is adapted from LOAM.

Contributors