FAST-LIO Then Bayesian ICP Then GTSFM Jerred Chen2 Xiangcheng Hu1 Shicong Ma2 Jianhao Jiao1 Ming Liu1 and Frank Dellaert2 Abstract For the Hilti Challenge 2022 we created two sys-

2025-04-27 0 0 2.15MB 4 页 10玖币

侵权投诉

FAST-LIO, Then Bayesian ICP, Then GTSFM

Jerred Chen2, Xiangcheng Hu1, Shicong Ma2, Jianhao Jiao1, Ming Liu1, and Frank Dellaert2

Abstract— For the Hilti Challenge 2022, we created two sys-

tems, one building upon the other. The ﬁrst system is FL2BIPS

which utilizes the iEKF algorithm FAST-LIO2 and Bayesian

ICP PoseSLAM, whereas the second system is GTSFM, a

structure from motion pipeline with factor graph backend

optimization powered by GTSAM.

I. OVERVIEW

We did two separate submissions and this report describes

both of them:

1) FL2BIPS: FAST-LIO2 + Bayesian ICP + Pose SLAM,

based on LIDAR and IMU only.

2) FL2BIPS-GTSFM The same system with addition-

ally using bundle adjustment, provided by GTSFM, a

GTSAM-backed Structure from Motion pipeline.

We start with FAST-LIO2 [1], a ﬁltering-based approach,

followed by a batch-optimization phase in which we re-

estimate the pose constraints and loop closures discovered

by FAST-LIO2, using Bayesian ICP [2]. The combination

of FAST-LIO2 and Bayesian ICP uses only LIDAR and

IMU, and is the system we used for our ﬁrst set of sub-

missions. The second set of submissions additionally runs

rig bundle-adjustment on top of the PoseSLAM graph, using

the GTSFM system.

We discuss both systems below.

II. FL2BIPS

A. FAST-LIO2 (FL2)

We employ the tightly-coupled iterated Kalman ﬁlter-

based LiDAR-inertial state estimator: FAST-LIO2 [1] to

estimate initial poses of sensors (in the IMU frame). Under

the ﬁlter-based framework: motion propagation and measure-

ment update, FAST-LIO2’s pipeline can be summarized as:

1) Forward propagation on motion upon each IMU mea-

surement at a high rate (i.e., 400 Hz).

2) LiDAR point de-skew: FAST-LIO2 designs the back-

ward propagation to estimate the LiDAR pose of each

point in the scan with respect to the pose at the scan

end time based on IMU measurements. With these

poses, each LiDAR point is transformed into the frame

at the end scanning time to correct the in-scan motion.

3) Iterated Kalman ﬁlter update: In each iteration, the de-

skewed LiDAR scan is matched with the global LiDAR

map by ﬁnding a set of point-to-plane correspondences.

1Robotics Institute, Department of Electronic and Computer Engineer-

ing, The Hong Kong University of Science and Technology {xhubd,

jjiao}@connect.ust.hk,{eelium}@ust.hk

2Institute for Robotics and Intelligent Machines, College

of Computing, Georgia Institute of Technology {jerred,

dellaert}@cc.gatech.edu

Based on the current updated state, FAST-LIO2 match

the current frame points with the map points, then

sensors’ motion are iteratively reﬁned by minimizing

the point-to-plane residuals. If the state converges,

FAST-LIO2 takes the state of the last iteration as the

posterior estimate.

4) With the estimated motion, the current LiDAR scan

is transformed into the global coordinate system and

incorporated with the global map. FAST-LIO2 also

proposes the iKD-Tree data structure, enabling incre-

mental point insertion and deletion as well as dynamic

rebalancing.

FAST-LIO2 is an online SLAM system. The iKD-Tree

design guarantees that the computation time of FAST-LIO2

is not affected by the environmental scale. On our desktop

computer: Intel i7-12700K, 20-thread CPU and 64GB RAM,

FAST-LIO2 takes around 15 to 30 ms to process each LiDAR

scan while simultaneously maintaining a global map. The

noise setting of the IMU is essential to the LIO system. Using

an open-source tool1, we calibrated the Allan variance given

the IMU calibration sequence provided by the organizer.

But the calibrated parameters did not boost the performance

of the LIO system. Thus, we directly use the IMU noise

and bias parameters from the calibration ﬁle. From our

experiments, FAST-LIO2 performs well on all sequences

except for Exp03,Exp09, and Exp15. These three sequences

present challenging scenarios for LiDAR-based systems: stair

and narrow corridor [3]. This motivates for utilizing the batch

estimation phase to further improve the pose estimates.

B. Bayesian ICP PoseSLAM (BIPS)

The batch estimation phase constructs a pose graph with

the Bayesian ICP pose constraints and perform a batch opti-

mization to reduce the drift error from the FAST-LIO2 pose

estimation. We utilize the Bayesian ICP algorithm to generate

pose constraints because Bayesian ICP can calculate the

relative transform between point clouds as well as estimate

the covariance online. To do this, Bayesian ICP samples the

posterior distribution of the relative transform using stochas-

tic gradient Langevin dynamics (SGLD) [4]. For this phase,

the pipeline contains two steps: pose constraint generation

and pose graph optimization. The following consist of major

facets of the pipeline.

1) Utilize a point-to-plane loss function for Bayesian ICP.

For each possible edge, the module reads in the FAST-

LIO2 poses output and use the FAST-LIO2 poses

1https://github.com/ori-drs/allan_variance_ros

arXiv:2210.00146v2 [cs.RO] 5 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

FAST-LIO,ThenBayesianICP,ThenGTSFMJerredChen2,XiangchengHu1,ShicongMa2,JianhaoJiao1,MingLiu1,andFrankDellaert2AbstractFortheHiltiChallenge2022,wecreatedtwosys-tems,onebuildingupontheother.TherstsystemisFL2BIPSwhichutilizestheiEKFalgorithmFAST-LIO2andBayesianICPPoseSLAM,whereasthesecondsystemisGTSF...

展开>> 收起<<

FAST-LIO Then Bayesian ICP Then GTSFM Jerred Chen2 Xiangcheng Hu1 Shicong Ma2 Jianhao Jiao1 Ming Liu1 and Frank Dellaert2 Abstract For the Hilti Challenge 2022 we created two sys-.pdf

共4页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

FAST-LIO Then Bayesian ICP Then GTSFM Jerred Chen2 Xiangcheng Hu1 Shicong Ma2 Jianhao Jiao1 Ming Liu1 and Frank Dellaert2 Abstract For the Hilti Challenge 2022 we created two sys-

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: