OpenLiDARMap: Zero-Drift Point Cloud Mapping Using Map Priors

Dominik Kulmer

† a

, Maximilian Leitenstern

, Marcel Weinmann

and Markus Lienkamp

Institute of Automotive Technology, Munich Institute of Robotics and Machine Intelligence,

Technical University of Munich, Garching, Germany

†

Keywords:

Mapping, Localization, SLAM, Georeferencing.

Abstract:

Accurate localization is a critical component of mobile autonomous systems, especially in Global Navigation

Satellite Systems (GNSS)-denied environments where traditional methods fail. In such scenarios, environ-

mental sensing is essential for reliable operation. However, approaches such as LiDAR odometry and Simul-

taneous Localization and Mapping (SLAM) suffer from drift over long distances, especially in the absence of

loop closures. Map-based localization offers a robust alternative, but the challenge lies in creating and georef-

erencing maps without GNSS support. To address this issue, we propose a method for creating georeferenced

maps without GNSS by using publicly available data, such as building footprints and surface models derived

from sparse aerial scans. Our approach integrates these data with onboard LiDAR scans to produce dense,

accurate, georeferenced 3D point cloud maps. By combining an Iterative Closest Point (ICP) scan-to-scan

and scan-to-map matching strategy, we achieve high local consistency without suffering from long-term drift.

Thus, we eliminate the reliance on GNSS for the creation of georeferenced maps. The results demonstrate that

LiDAR-only mapping can produce accurate georeferenced point cloud maps when augmented with existing

map priors.

1 INTRODUCTION

Localization is essential for mobile autonomous sys-

tems, enabling them to navigate and interact with

their environment effectively. In recent years, signiﬁ-

cant advancements have been made in LiDAR odom-

etry (Vizzo et al., 2023; Zheng and Zhu, 2024b),

LiDAR-inertial odometry (Bai et al., 2022; Xu et al.,

2022; Zheng and Zhu, 2024a), and Simultaneous Lo-

calization and Mapping (SLAM) (Dellenbach et al.,

2022; Yifan et al., 2024; Koide et al., 2024b; Pan

et al., 2024). Despite their advances, these methods

remain prone to drift, especially during long-term op-

erations. While loop closure can help reduce drift by

aligning the current position with a previously visited

location, it does not guarantee an accurate reconstruc-

tion of the intermediate path. Additionally, loops are

not always present, especially in linear or open-ended

trajectories.

Map-based localization offers an alternative by us-

ing pre-existing maps to constrain localization and

https://orcid.org/0000-0001-7886-7550

https://orcid.org/0009-0008-6436-7967

https://orcid.org/0009-0008-7174-4732

https://orcid.org/0000-0002-9263-5323

Figure 1: Representation of the different map priors and

formats from the building approximations on the left to the

ﬁnal georeferenced point cloud map on the right for KITTI

Seq. 00.

eliminate drift. However, these algorithms rely on the

availability of maps, which limits their applicability

in unknown environments where LiDAR odometry,

LiDAR-inertial odometry, and SLAM excel. Open-

178

Kulmer, D., Leitenstern, M., Weinmann, M. and Lienkamp, M.

OpenLiDARMap: Zero-Drift Point Cloud Mapping Using Map Priors.

DOI: 10.5220/0013405400003941

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 11th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2025), pages 178-188

ISBN: 978-989-758-745-0; ISSN: 2184-495X

StreetMap

(OSM), a globally available web-based

platform, provides maps that can be used for localiza-

tion with minimal geographic constraints. However,

these standard-deﬁnition (SD) maps cannot support

high-precision localization due to their limited accu-

racy.

High-deﬁnition (HD) maps represent the environ-

ment at a resolution of 10 cm to 20 cm (Jeong et al.,

2022) and are ideal for precise localization (Koide

et al., 2019; Koide et al., 2024a). These maps typ-

ically consist of high accuracy point cloud maps.

However, the map generation process suffers from

the same drift and accuracy issues as the localization

itself. Global Navigation Satellite Systems (GNSS)

offer a global reference to mitigate drift and enable

georeferenced mapping. However, their signals are

highly susceptible to degradation in obstructed en-

vironments. This vulnerability signiﬁcantly reduces

their reliability for producing high-precision maps.

This work addresses these challenges by propos-

ing a method for high-deﬁnition map generation that

combines publicly available building maps and sur-

face models with onboard LiDAR data. By leverag-

ing these resources, the approach eliminates reliance

on GNSS and enables the creation of accurate, geo-

referenced point cloud maps (Figure 1).

The main contribution of this paper is a pose-

graph-based optimization algorithm that combines It-

erative Closest Point (ICP) scan-to-scan with scan-to-

map matching of publicly available building data and

sparse surface models to generate accurate georefer-

enced point cloud maps without the need for GNSS

data. While scan-to-map matching regulates long-

term drift, scan-to-scan matching maintains local con-

sistency between individual LiDAR scans. It also en-

ables the approach to compensate for missing build-

ings and outdated map data or to bridge short areas

not present on the sparse prior map.

Our approach uses neither learning-based meth-

ods, feature extraction techniques, nor loop closures.

We show that the approach works with a single pa-

rameter set on different platforms, like vehicles and

segways, with different LiDAR setups, ranging from

a single 32-channel LiDAR to a modern multi-LiDAR

setup, and in various environments, such as residen-

tial and rural regions.

In sum, we make four claims: Our approach is

able to (I) map long sequences without accumulat-

ing drift over time; (II) automatically georeference the

generated map without GNSS data; (III) keep a high

local consistency of the generated map; (IV) yield

promising results on multiple robotic platforms, Li-

DAR setups and environments without further tuning.

www.openstreetmap.org

To build on our work, we pro-

vide an open-source implementation at:

https://github.com/TUMFTM/OpenLiDARMap

2 RELATED WORK

When available, high-precision RTK-GNSS signals

are often used for georeferenced mapping, requiring

precise time synchronization between sensors and op-

timal signal reception along the mapped route. How-

ever, GNSS signal interruptions, such as in urban

canyons or underpasses, pose a challenge to con-

tinuous mapping. Hybrid pipelines have been pro-

posed to deal with signal interruption. These typ-

ically involve creating an initial map using LiDAR

odometry or SLAM, followed by post-processing to

georeference the map using GNSS data (Leitenstern

et al., 2024). While this approach can bridge short

gaps without GNSS, it remains dependent on inter-

mittent signal availability. SLAM techniques fused

with GNSS signals can also produce georeferenced

maps (Koide et al., 2019; Cramariuc et al., 2023;

Dellaert, 2022), but require careful tuning of sen-

sor weights and suffer from the same dependency on

GNSS signal quality.

OSM represents an alternative to GNSS for global

localization and mapping, relying on widely avail-

able crowdsourced map data. (Floros et al., 2013)

improved global localization accuracy by combining

visual odometry with Monte Carlo localization, us-

ing chamfer matching to align trajectories with OSM

maps. (Suger and Burgard, 2017) proposed a proba-

bilistic navigation method that aligns 3D-LiDAR sen-

sor data with OSM tracks using semantic terrain infor-

mation and a Markov-Chain Monte Carlo framework.

(Yang et al., 2017) introduced a Gaussian-Gaussian

cloud model for visual odometry, where OSM road

constraints help mitigate drift and resolve scale am-

biguities. (Yan et al., 2019) used OSM to create

orthophoto-style images of roads and building foot-

prints and generated semantic descriptors to match

LiDAR data. (Ballardini et al., 2021) introduced a lo-

calization method that detects building facades using

stereo image-based point clouds and matches them

with 3D building models from OSM. Similarly, (Cho

et al., 2022) computed angular distances to buildings

within OSM and produced descriptors that match Li-

DAR data to achieve localization. (Elhousni et al.,

2022) used a particle ﬁlter to integrate LiDAR point

clouds with OSM constraints such as road bound-

aries, improving accuracy by exploiting map geom-

etry. (Frosi et al., 2023) extended this concept by

combining SLAM with OSM priors, integrating 2D

OpenLiDARMap: Zero-Drift Point Cloud Mapping Using Map Priors

179

building geometry into trajectory estimation through

LiDAR scan matching.

These methods highlight the diverse applications

of OSM in reducing drift for mapping and localiza-

tion in the absence of GNSS signals. However, chal-

lenges such as limited accuracy and dependence on

map accuracy remain.

Satellite or aerial imagery is another alternative

for georeferenced mapping by aligning ground-level

observations with overhead map features. (Miller

et al., 2021) presented a cross-view localization

framework that leverages semantic LiDAR point

clouds alongside top-down RGB satellite imagery for

georeferenced mapping. Similarly, (Xia et al., 2024)

reﬁned cross-view localization techniques to operate

effectively in areas lacking ﬁne-grained ground truth,

relying on coarse satellite or map data to improve ac-

curacy. However, the inherent misalignment between

ground observations and aerial imagery remains a per-

sistent challenge, often resulting in signiﬁcant local-

ization errors.

Pure LiDAR odometry and SLAM methods lack

georeferencing, while GNSS-based approaches re-

main vulnerable to environmental interference and

signal loss. OSM-based methods, while independent

of GNSS, rely on the accuracy of crowd-sourced maps

that may be outdated or inaccurate. Satellite data pro-

vides broad coverage but often suffers from alignment

problems between ground and aerial imagery.

These limitations highlight the need for novel

approaches that can achieve georeferenced mapping

without relying on GNSS.

3 GEOREFENCED POINT

CLOUD MAPPING

In this section, we explain our LiDAR-based point

cloud mapping approach. The main idea is to com-

bine scan-to-scan matching with scan-to-map match-

ing of reference maps of publicly available build-

ing information and sparse surface models in a pose-

graph optimization framework. The advantage of our

method is that we maintain high local consistency

while eliminating long-term drift without the need for

additional sensors. Our mapping procedure can be

summarized in ﬁve steps (Figure 3); while the ﬁrst

one is done beforehand, the others are performed for

each input frame:

1. Generating a sparse reference point cloud map

from openly available building data and surface

models.

2. Scan matching of the current LiDAR scan and the

Figure 2: Exemplary illustration of a combined sparse

point cloud map of (green) approximated building data and

(black) the surface model with a ground sampling distance

of 1 m around the starting position of KITTI Seq. 00.

sparse reference map.

3. Scan matching of the current LiDAR scan and a

local submap of previous LiDAR scans.

4. Performing a graph optimization of the resulting

poses.

5. Estimating the initial guess for the next LiDAR

scan from a constant distance and rotation as-

sumption.

3.1 Generating Sparse Reference Maps

OSM provides, besides road infrastructure data,

building outlines for much of the world. In addi-

tion to geometric outlines, OSM can store semantic

attributes such as building height or number of sto-

ries. Beyond OSM, other publicly available local data

sources, such as Germany’s open data portal

, provide

detailed building information. From this data, sparse

georeferenced three-dimensional point clouds can be

generated. Depending on data availability, a simpli-

ﬁed building representation is derived from OSM data

or more sophisticated building models, such as those

from the German open data portal. While building

models can directly be approximated as a point cloud

due to their inherent spatial representation, OSM data

requires an estimation for building heights. We as-

sume a height of 4 m per ﬂoor, with a default height of

8 m for untagged buildings. The resulting point cloud

approximates the building shape using a tessellation

with an edge length of 0.5 m.

In addition to building data, three-dimensional

surface measurements are available for many regions

of the world

2345

. These surface measurements are

www.govdata.de

www.data.europa.eu

www.data.gov.uk

www.usgs.gov

VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems

180

Figure 3: Pipeline overview that shows our main steps to generate georeferenced point cloud maps. Starting with (1) the

sparse reference map, which is used for (2) the scan-to-map matching. (3) shows the scan-to-scan matching. The two results

are then optimized with (4) a pose-graph optimization, resulting in the ﬁnal pose, which is then used for (5) the initial guess

estimation for the next LiDAR frame.

typically provided as either surface models or eleva-

tion models with varying ground sampling distances,

often in the region of 1 m to 5 m. Surface models

capture raw LiDAR scans or a downsampled subset,

while elevation models represent the terrain alone, ex-

cluding vegetation and structures. This data is just an-

other representation of point clouds. Therefore, it can

be combined with the building point cloud data to cre-

ate a sparse georeferenced point cloud map (Figure 2).

We use this map as a reference in the following scan-

to-map matching.

3.2 Scan-to-Map Matching

Our approach is based on small gicp (Koide, 2024),

a lightweight, header-only C++ library designed for

point cloud preprocessing and scan matching. We

adopt a voxel hash map as our data structure to

efﬁciently handle large point clouds, allowing fast

nearest-neighbor searches. The data within each

voxel is stored according to the linear iVox princi-

ple (Bai et al., 2022).

The current input frame is pre-processed with a

voxel-based downsampling before the scan matching.

The parameters for our entire processing chain can be

taken from subsection 3.6.

A manual initial pose estimate on the map is pro-

vided for the ﬁrst input frame and subsequently re-

ﬁned through scan-to-map matching. For subsequent

frames, the initial estimate is derived from the re-

sults of the pose-graph optimization, as described in

subsection 3.5. The scan matching process uses the

ICP algorithm for its simplicity and robustness. Out-

liers are managed using the Geman-McClure robust

kernel, and the optimization is carried out using the

Gauss-Newton method. The scan matching process

returns a global pose within our reference map.

3.3 Scan-to-Scan Matching

The same procedure used for scan-to-map matching

is applied to scan-to-scan matching, using the same

pre-processing, data structure, and optimization.

For scan-to-scan matching, a local submap of the

environment is incrementally constructed. The ﬁrst

input frame serves as the registration target for the

subsequent scan frame. If the relative transformation

between consecutive frames is below 0.1 m, the sys-

tem assumes a static state. Therefore, no pose is pro-

vided for the optimization, and the optimization pro-

cess for both scan-to-scan and scan-to-map matching

is skipped. In dynamic scenarios, as deﬁned by this

threshold, the scan-to-scan matching outputs the rela-

tive transformation.

After the pose-graph optimization, the current

frame is integrated into the local submap using the

determined transformation. To maintain a local rep-

resentation of the environment, any voxels in the

submap that are more than 100 m from the current po-

sition are removed. From the second frame onward,

OpenLiDARMap: Zero-Drift Point Cloud Mapping Using Map Priors

181

scan matching is performed against this evolving local

submap rather than solely against the previous frame.

This approach is inspired by KISS-ICP (Vizzo et al.,

2023) and leverages the initial pose estimation de-

scribed in subsection 3.5.

3.4 Pose-Graph Optimization

Our approach integrates scan-to-scan and scan-to-

map matching into an optimization framework using

the Ceres solver (Agarwal and Mierle, 2023). The

optimization problem is formulated as a non-linear

least squares problem (Equation 1), where residuals

are minimized to align the current pose estimate with

the scan matching results.

min

∑



|| f

, ·· · , x

)||



(1)

represents a loss function to reduce the in-

ﬂuence of outliers on the solution. f

( ) is the

cost function, which depends on the parameter block

, ·· · , x

Unlike batch optimization methods, which pro-

cess all frames simultaneously, the proposed ap-

proach employs frame-by-frame optimization. For

each frame, the residuals are computed as the devi-

ation between the initial pose estimate and the results

obtained from scan-to-scan and scan-to-map match-

ing. Scan-to-map constraints represent absolute pose

estimates derived from aligning the current frame

to the sparse map. Scan-to-scan constraints rep-

resent relative transformations between consecutive

frames, linking absolute poses through relative mea-

surements. These assemble a classic pose-graph opti-

mization problem.

To mitigate the inﬂuence of outliers, residuals are

weighted using a robust loss function to reduce the

impact of large deviations on the ﬁnal solution. We

use a Tukey loss to aggressively suppress large devia-

tions for the scan-to-map constraints. For the scan-to-

scan matching, we use the softer Cauchy loss. These

weights reduce the inﬂuence of outliers, which are

particularly likely in scan-to-map matching due to de-

viations between the sparse reference map and the

current LiDAR frame. To maintain the reliability of

scan-to-map constraints, they are only incorporated

into the optimization process if the number of point

inliers exceeds 50 % of the correspondences between

the current frame and the sparse map. This crite-

rion ensures that only sufﬁciently aligned frames con-

tribute to the optimization, mitigating the effect of

poor matches.

The optimization framework balances relative

constraints from scan-to-scan matching with absolute

constraints from scan-to-map matching, providing ac-

curate and drift-corrected pose estimates. The result

of the optimization is a georeferenced pose computed

from the results of the scan-to-scan and scan-to-map

matching.

3.5 Estimating the Initial Guess

We use a constant distance and rotation model for the

initial pose estimation. The delta translation between

the previous and current position is applied to the cur-

rent position to estimate the next initial position.

x+1

= t

+ (t

− t

x−1

) (2)

For the rotation we use a spherical quaternion ex-

trapolation through slerp (Equation 3) with a scalar of

s = 2.

sler p(q

x−1

, q

, s) = q

x−1

−1

x−1

)

(3)

This straightforward method eliminates the need

for time stamping across point clouds or individual

points, simplifying the implementation. In addition,

the lack of a complex kinematic model enhances the

versatility of the approach, allowing seamless appli-

cation across different robotic platforms without the

need to modify the methodology.

3.6 Parameters

We use a single set of parameters for both scan-to-

scan and scan-to-map matching, as well as for all sub-

sequent evaluations. This design choice is aimed at

simplicity, minimizing the number of parameters and

allowing for a more straightforward system conﬁgu-

ration.

Each input frame is downsampled with a 1.5 m

voxel ﬁlter to reduce the computational load. The

voxel maps are constructed with a resolution of 1 m

and can store up to 10 points per voxel, maintain-

ing a minimum distance of 0.1 m between individual

points.

For the nearest neighbor search, we consider 27

neighboring voxels within a 3x3x3 cube around the

query point. A correspondence threshold of 6 m is

applied to ﬁlter out invalid associations during scan

matching. For all robust kernels used in the opti-

mization process, we use a static kernel width of 1.0

to ensure consistent outlier handling throughout the

pipeline.

4 EXPERIMENTAL EVALUATION

We present our experiments to show the capabilities

of our method. The results of our experiments sup-

VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems

182

port our key claims, namely that we can (I) map long

sequences without accumulating drift over time; (II)

automatically georeference the generated map with-

out GNSS data; (III) keep a high local consistency

of the generated map; (IV) yield promising results on

multiple robotic platforms, LiDAR setups and envi-

ronments without further tuning.

4.1 Experimental Setup

We evaluate our method on the three different long-

distance datasets KITTI Seq. 00 (Geiger et al., 2012;

Geiger et al., 2013), NCLT 2013-01-10 (Carlevaris-

Bianco et al., 2016), and EDGAR Campus, each us-

ing different robot platforms and sensor setups. The

characteristics can be found in Table 1. The KITTI

dataset is a widely accepted benchmark for evalu-

ating localization and mapping algorithms. Follow-

ing the approach demonstrated by IMLS-SLAM (De-

schaud, 2018) and later adopted by others, such as

KISS-ICP (Vizzo et al., 2023) and CT-ICP (Dellen-

bach et al., 2022), we apply a correction to the intrin-

sic calibrations around the vertical axis. Speciﬁcally,

the point clouds are adjusted by an angle of 0.375

◦

ensure accurate alignment.

The NCLT dataset employs a Segway platform,

introducing signiﬁcant challenges due to the vehi-

cle’s high dynamics and diverse environmental condi-

tions. The dataset encompasses outdoor scenes rang-

ing from wide-open parking areas to narrow building

passages.

The EDGAR dataset is the longest of the three

tested datasets and covers the Technical University of

Munich campus in Garching, Germany. It includes a

variety of environments, from building complexes and

extensive parking areas to country roads, with vehi-

cle speeds ranging from low speeds up to 70 kmh

−1

The route traverses both dense and sparsely built-up

areas. Unlike KITTI and NCLT, EDGAR features a

multi-LiDAR setup, consisting of two spinning and

two solid-state LiDARs, which enhances data rich-

ness and coverage. While the ﬁrst two datasets are

freely available, the EDGAR Campus dataset is pro-

prietary. Information about the sensor setup can be

found in (Karle et al., 2023; Kulmer et al., 2024).

All vehicles are equipped with RTK-GNSS sys-

tems to provide the ground truth data.

The quality of the maps created cannot be an-

swered with a single metric. A distinction is made

between global deviations, such as drift, and local in-

consistencies.

Table 1: Caracteristic of the datasets used for the evaluation.

Dataset

Length

[m]

Frames

[#]

Sensor

Setup

Scenario

KITTI

Seq. 00

3724 4541 Velodyne HDL-64 outdoor, residential, car

NCLT

2013-01-1

1311 5120 Velodyne HDL-32E outdoor/indoor, residential, segway

EDGAR

Campus

5289 9104

2x Ouster OS1-128

2x Seyond Falcon

outdoor, rual/residential, car

4.2 Global Displacement Evaluation

We evaluate our approach using the Absolute Tra-

jectory Error (ATE) and the KITTI metric for the

Relative Trajectory Error (RTE). The KITTI met-

ric samples the trajectory into segments of varying

lengths, ranging from 100 m to 800 m, and calcu-

lates the average relative translational and rotational

errors. While the ATE measures the global deviation

of the estimated trajectory from the ground truth, the

KITTI metric provides insights into the accuracy over

medium-length segment pieces.

We compare our approach with three OSM-based

global localization and mapping methods and a satel-

lite imagery-based approach for the ATE evalua-

tion. Additionally, we include KISS-ICP as a LiDAR

odometry baseline for comparison. However, open-

source code is unavailable for all three of the OSM-

based methods, and the repository for the imagery-

based method is outdated. As a result, we rely on the

published data from their respective published works.

To enable a comparison with the ground truth, which

is positioned in a local coordinate system originating

in the ﬁrst frame, we transform our georeferenced re-

sults into the ﬁrst frame to obtain the same local repre-

sentation. Then, we calculate the mean and maximum

ATE using the ground truth data with evo

The results listed in Table 2 demonstrate that our

approach reduces the mean ATE by more than half

compared to the next best method. Furthermore, our

method achieves a mean ATE that is a magnitude

lower than the LiDAR odometry approach. Figure 4

shows that our method accumulates no drift over the

traversed distance. Notably, we observe an unusually

high deviation in a speciﬁc section compared to the

rest of the trajectory. Figure 4 (I) and (II) show the

point cloud generated from the ground truth data and

our approach for this section. We identify signiﬁcant

deviations in the provided ”ground truth” data for the

two passes of the section, particularly with vertical

offsets exceeding 1 m in some regions. Hence, the

ground truth provided by KITTI cannot be regarded as

accurate values, affecting the overall evaluation and

the metrics presented.

https://github.com/MichaelGrupp/evo

OpenLiDARMap: Zero-Drift Point Cloud Mapping Using Map Priors

183

Figure 4: Absolute Trajectory Error (ATE) of our approach for KITTI Seq. 00. (I) Point cloud created from the ”ground truth”

data. (II) Point cloud created with our approach.

Figure 5: Absolute Trajectory Error (ATE) of our approach

for the EDGAR Campus dataset.

Figure 6: Absolute Trajectory Error (ATE) of our approach

for the NCLT 2013-01-1 dataset.

Beyond the KITTI dataset, we demonstrate simi-

lar performance for the NCLT and EDGAR datasets,

where our method exhibits no drift over time. On the

EDGAR dataset, again the largest ATEs occur in the

section with bad GNSS signals, most likely due to the

signal blockage caused by underpasses and tall build-

ing walls (Figure 5). For NCLT, although deviations

are the largest, our pose estimation remains robust de-

spite the challenging dynamics and even across an un-

mapped indoor section (Figure 6).

Figure 7: KITTI Seq. 00 point cloud map, created with our

approach, plotted to the orthophoto of Karlsruhe.

We obtain a global pose directly from our ap-

proach and only transform it into a local coordinate

system for evaluation. On Figure 7, we show the over-

load of the directly generated map by our approach on

georeferenced orthophotos of Karlsruhe

In addition to the ATE, we also evaluated the

RTE using the KITTI methodology (Table 3). For

the KITTI dataset, our approach is outperformed by

the LiDAR odometry algorithm KISS-ICP. This result

can be attributed to the high consistency and smooth

trajectory generated by KISS-ICP, which minimizes

drift over the short averaged segments used in the

KITTI metric. However, on the NCLT dataset, which

features high dynamics and a challenging environ-

ment, our method demonstrates a better performance.

VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems

184

Table 2: Mean and max Absolute Trajectory Error (ATE) in [m] for the evaluated datasets. Bold represents the best results

and underscores the second-best results.

Method Reference

KITTI Seq. 00 NCLT 2013-01-1 EDGAR Campus

mean ↓ max ↓ mean ↓ max ↓ mean ↓ max ↓

RC-MVO (Yang et al., 2017) 3.76 14.01 - - - -

AWYLaI (Miller et al., 2021) 2.0 12.0 - - - -

LiDAR-OSM (Elhousni et al., 2022) 1.37 3.34 - - - -

OSM-SLAM (Frosi et al., 2023) 3.15 11.06 - - - -

KISS-ICP (Vizzo et al., 2023) 6.71 15.20 3.26 15.99 66.40 259.13

OpenLiDARMap (ours) 0.66 2.19 1.02 3.81 0.52 2.84

Similarly, the advantages of our approach are evident

on the EDGAR dataset, where we maintain consistent

trajectories over long distances without drift, leading

to lower relative errors.

To assess how well our approach per-

forms compared to pure scan-to-map algo-

rithms, we tried to evaluate the current state

of the art on our sparse reference map. How-

ever, none of LiLoc (Fang et al., 2024), BM-

Loc (Feng et al., 2024), HDL localization (Koide

et al., 2019), DLL (Caballero and Merino,

2021), FAST LIO LOCALIZATION

, and li-

dar localization ros2

was able to ﬁnish KITTI Seq.

00, leaving us without any results to compare.

4.3 Local Consistency Evaluation

For quantitative evaluation of map consistency, we

use the Mean Map Entropy (MME) metric (Razlaw

et al., 2015). Speciﬁcally, we compare the MME

between the point cloud map generated by OpenL-

iDARMap and the map generated from the ground

truth transformations of the individual scans.

To facilitate this comparison, we preprocess the

point clouds by cropping them to a 100×100 m re-

gion and downsampling individual clouds to a voxel

size of 0.25 m. Notably, the composite map remains

unﬁltered to avoid introducing bias into the evalua-

tion. This preprocessing step is essential to manage

the computational complexity caused by the size of

the maps.

In the case of the KITTI dataset, we can also use

the MME to show that the ground truth is not exact.

Our map shows a better (lower) score compared to the

ground truth (Table 4). The same can also be shown

for the NCLT results, where our approach reaches a

https://github.com/HViktorTsoi/fast lio localization

https://github.com/rsasaki0109/lidar localization ros2

lower score than the map generated from the ground

truth data. For the EDGAR dataset, the MME gener-

ated with the ground truth is lower than the results of

our approach.

4.4 Out-of-Date Reference Maps

Finally, we demonstrate that our method does not rely

on up-to-date reference maps. Across all datasets,

we show that neither current map data nor temporal

alignment between the maps and the LiDAR frames

is necessary. For instance, the surface data used

in the KITTI dataset originates from 2000 to 2023,

while the building data stems from 2023, and the

evaluated dataset was collected in 2011. Similarly,

the NCLT dataset’s data sources span from 2013 to

2024, whereas the EDGAR dataset spans from 2023

to 2024.

The accompanying Figure 8 and Figure 9 high-

light discrepancies between the sparse maps and the

datasets, such as missing buildings or structural dif-

ferences. These variations emphasize the minimal re-

quirements for input data in our approach. For in-

stance, building outlines for the NCLT dataset were

extracted from OSM data, whereas those for KITTI

and EDGAR were derived from spatial building mod-

els. Moreover, the resolution and type of input

data vary signiﬁcantly: KITTI and NCLT datasets

rely on 1 m-resolution GeoTIFFs, while the EDGAR

dataset beneﬁts from raw LiDAR point clouds ob-

tained through aerial scans. This diversity highlights

the adaptability of our method to sparse maps with

signiﬁcantly different compositions.

As a ﬁnal note, our approach focuses on mapping,

making runtime not a concern for us. Despite that, we

are able to run the entire pipeline in about 30 ms per

KITTI frame on a modern PC with an AMD Ryzen

7700.

OpenLiDARMap: Zero-Drift Point Cloud Mapping Using Map Priors

185

Table 3: Relative Trajectory Error (RTE) using the KITTI methodology for the evaluated datasets.

Method Reference

KITTI Seq. 00 NCLT 2013-01-1 EDGAR Campus

trans. ↓

[%]

rot. ↓

[deg/m]

trans. ↓

[%]

rot. ↓

[deg/m]

trans. ↓

[%]

rot. ↓

[deg/m]

KISS-ICP (Vizzo et al., 2023) 0.51 0.0017 2.31 0.0161 2.77 0.0095

OpenLiDARMap (ours) 0.53 0.0025 1.93 0.0116 1.44 0.0046

Table 4: Mean Map Entropy for the frames transformed

with the ground truth data and OpenLiDARMap results.

Lower ↓ is better.

Method

KITTI

Seq. 00

NCLT

2013-01-1

EDGAR

Campus

Ground Truth -6.293 -6.212 -6.317

OpenLiDARMap (ours) -6.448 -6.339 -6.231

Figure 8: Overload of the sparse reference map and the ﬁnal

point cloud map created with our approach for the NCLT

2013-01-1 dataset. (I) shows part of the reference map, and

(II) shows the point cloud map generated with our approach

for an indoor section of the dataset that we were able to map

despite the missing correspondences between the onboard

LiDAR frames and the reference map.

5 LIMITATIONS AND

FUTURE WORK

Unlike the ATE analysis, where our approach showed

clear improvements, the RTE analysis yielded less

conclusive results. This result can be primarily at-

Figure 9: Overload of the sparse reference map and the ﬁnal

point cloud map created with our approach for the EDGAR

Campus dataset. Highlighted is a part of the map that still

shows (I) the start of construction with multiple cranes in

the reference map, while (II) the generated map from the

LiDAR frames shows the completed building facades.

tributed to rotational errors, which have a particu-

larly strong impact on the KITTI metric. Future work

could address this limitation by improving the initial

estimation, such as using homogeneous transforms or

incorporating the time differences between frames, as

well as using Lie Algebra in the optimization to better

account for small displacements.

A key requirement for the functionality of our ap-

proach is the availability of a digital surface model.

While such data is available for large parts of the

world, it is not universally accessible. A notable ex-

ample is South Korea, where the lack of digital sur-

face models prevents us from evaluating our approach

on widely used datasets such as MulRan (Kim et al.,

2020) or HeLiPR (Jung et al., 2024).

Due to the simple ICP-based scan-to-scan match-

ing strategy, our method is limited in handling un-

mapped regions, such as long tunnels or extensive

indoor areas. However, the modular design of our

approach allows for future integration of more ad-

vanced LiDAR Odometry algorithms to address these

challenges. Additionally, signiﬁcant environmental

VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems

186

changes can lead to erroneous results, emphasizing

the need for further investigations across diverse sce-

narios to evaluate robustness under varying condi-

tions.

6 CONCLUSIONS

In this paper, we presented an approach for geo-

referenced point cloud mapping that operates with-

out the need for GNSS. By combining a scan-to-

map and scan-to-scan point cloud registration method

within an optimization framework, we achieved accu-

rate mapping across diverse environments, vehicles,

and LiDAR setups. Our experiments suggest that

the proposed method effectively eliminates drift over

long distances and is robust to variations in environ-

mental conditions and sensor conﬁgurations. Our ex-

periments suggest that the proposed method enables

accurate georeferenced point cloud mapping without

relying on GNSS sensors.

ACKNOWLEDGEMENTS

As the ﬁrst author, Dominik Kulmer initiated and de-

signed the paper’s structure. He is the main contrib-

utor to the design and implementation of the overall

concept. Maximilian Leitenstern, Marcel Weinmann,

and Markus Lienkamp revised the paper critically

for important intellectual content. Markus Lienkamp

gives ﬁnal approval for the version to be published

and agrees to all aspects of the work. As a guaran-

tor, he accepts responsibility for the overall integrity

of the paper.

The research was partially funded by the Bavar-

ian Research Foundation (BFS) and through basic re-

search funds from the Institute of Automotive Tech-

nology (FTM).

REFERENCES

Agarwal, S. and Mierle, K. (2023). Ceres solver.

https://github.com/ceres-solver/ceres-solver.

Bai, C., Xiao, T., Chen, Y., Wang, H., Zhang, F., and Gao,

X. (2022). Faster-lio: Lightweight tightly coupled

lidar-inertial odometry using parallel sparse incremen-

tal voxels. IEEE Robotics and Automation Letters,

7:4861–4868.

Ballardini, A. L., Fontana, S., Cattaneo, D., Matteucci, M.,

and Sorrenti, D. G. (2021). Vehicle localization using

3d building models and point cloud matching. Sen-

sors, 21:5356.

Caballero, F. and Merino, L. (2021). Dll: Direct lidar

localization. a map-based localization approach for

aerial robots. In 2021 IEEE/RSJ International Confer-

ence on Intelligent Robots and Systems (IROS), pages

5491–5498. IEEE.

Carlevaris-Bianco, N., Ushani, A. K., and Eustice, R. M.

(2016). University of michigan north campus long-

term vision and lidar dataset. The International Jour-

nal of Robotics Research, 35:1023–1035.

Cho, Y., Kim, G., Lee, S., and Ryu, J.-H. (2022).

Openstreetmap-based lidar global localization in ur-

ban environment without a prior lidar map. IEEE

Robotics and Automation Letters, 7:4999–5006.

Cramariuc, A., Bernreiter, L., Tschopp, F., Fehr, M., Rei-

jgwart, V., Nieto, J., Siegwart, R., and Cadena, C.

(2023). maplab 2.0 – a modular and multi-modal map-

ping framework. IEEE Robotics and Automation Let-

ters, 8:520–527.

Dellaert, F. (2022). borglab/gtsam.

https://github.com/borglab/gtsam, 10.5281/zen-

odo.5794541.

Dellenbach, P., Deschaud, J.-E., Jacquet, B., and Goulette,

F. (2022). Ct-icp: Real-time elastic lidar odometry

with loop closure. In 2022 International Conference

on Robotics and Automation (ICRA), pages 5580–

5586. IEEE.

Deschaud, J.-E. (2018). Imls-slam: Scan-to-model match-

ing based on 3d data. In 2018 IEEE International

Conference on Robotics and Automation (ICRA),

pages 2480–2485. IEEE.

Elhousni, M., Zhang, Z., and Huang, X. (2022). Lidar-osm-

based vehicle localization in gps-denied environments

by using constrained particle ﬁlter. Sensors, 22:5206.

Fang, Y., Li, Y., Qian, K., Tombari, F., Wang, Y., and Lee,

G. H. (2024). Liloc: Lifelong localization using adap-

tive submap joining and egocentric factor graph.

Feng, Y., Jiang, Z., Shi, Y., Feng, Y., Chen, X., Zhao,

H., and Zhou, G. (2024). Block-map-based localiza-

tion in large-scale environment. In 2024 IEEE In-

ternational Conference on Robotics and Automation

(ICRA), pages 1709–1715. IEEE.

Floros, G., van der Zander, B., and Leibe, B. (2013). Open-

streetslam: Global vehicle localization using open-

streetmaps. In 2013 IEEE International Confer-

ence on Robotics and Automation, pages 1054–1059.

IEEE.

Frosi, M., Gobbi, V., and Matteucci, M. (2023). Osm-slam:

Aiding slam with openstreetmaps priors. Frontiers in

Robotics and AI, 10.

Geiger, A., Lenz, P., Stiller, C., and Urtasun, R. (2013).

Vision meets robotics: The kitti dataset. The Interna-

tional Journal of Robotics Research, 32:1231–1237.

Geiger, A., Lenz, P., and Urtasun, R. (2012). Are we ready

for autonomous driving? the kitti vision benchmark

suite. In 2012 IEEE Conference on Computer Vision

and Pattern Recognition, pages 3354–3361. IEEE.

Jeong, J., Yoon, J. Y., Lee, H., Darweesh, H., and Sung,

W. (2022). Tutorial on high-deﬁnition map generation

for automated driving in urban environments. Sensors,

22:7056.

OpenLiDARMap: Zero-Drift Point Cloud Mapping Using Map Priors

187

Jung, M., Yang, W., Lee, D., Gil, H., Kim, G., and Kim, A.

(2024). Helipr: Heterogeneous lidar dataset for inter-

lidar place recognition under spatiotemporal varia-

tions. The International Journal of Robotics Research,

43:1867–1883.

Karle, P., Betz, T., Bosk, M., Fent, F., Gehrke, N.,

Geisslinger, M., Gressenbuch, L., Hafemann, P., Hu-

ber, S., H

ubner, M., Huch, S., Kaljavesi, G., Kerbl,

T., Kulmer, D., Maierhofer, S., Mascetta, T., Pfab,

F., Rezabek, F., Rivera, E., Sagmeister, S., Seidlitz,

L., Sauerbeck, F., Tahiraj, I., Trauth, R., Uhlemann,

N., W

ursching, G., Zarrouki, B., Althoff, M., Betz,

J., Bengler, K., Carle, G., Diermeyer, F., Ott, J., and

Lienkamp, M. (2023). Edgar: An autonomous driv-

ing research platform – from feature development to

real-world application.

Kim, G., Park, Y. S., Cho, Y., Jeong, J., and Kim, A.

(2020). Mulran: Multimodal range dataset for urban

place recognition. In 2020 IEEE International Con-

ference on Robotics and Automation (ICRA), pages

6246–6253. IEEE.

Koide, K. (2024). small gicp: Efﬁcient and parallel algo-

rithms for point cloud registration. Journal of Open

Source Software, 9:6948.

Koide, K., Miura, J., and Menegatti, E. (2019). A portable

three-dimensional lidar-based system for long-term

and wide-area people behavior measurement. Inter-

national Journal of Advanced Robotic Systems, 16.

Koide, K., Oishi, S., Yokozuka, M., and Banno, A. (2024a).

Tightly coupled range inertial localization on a 3d

prior map based on sliding window factor graph opti-

mization. In 2024 IEEE International Conference on

Robotics and Automation (ICRA), pages 1745–1751.

IEEE.

Koide, K., Yokozuka, M., Oishi, S., and Banno, A. (2024b).

Glim: 3d range-inertial localization and mapping with

gpu-accelerated scan matching factors. Robotics and

Autonomous Systems, 179:104750.

Kulmer, D., Tahiraj, I., Chumak, A., and Lienkamp, M.

(2024). Multi-lica: A motion- and targetless multi

- lidar-to-lidar calibration framework. In 2024 IEEE

International Conference on Multisensor Fusion and

Integration for Intelligent Systems (MFI), pages 1–7.

IEEE.

Leitenstern, M., Sauerbeck, F., Kulmer, D., and Betz, J.

(2024). Flexmap fusion: Georeferencing and auto-

mated conﬂation of hd maps with openstreetmap.

Miller, I. D., Cowley, A., Konkimalla, R., Shivakumar,

S. S., Nguyen, T., Smith, T., Taylor, C. J., and Ku-

mar, V. (2021). Any way you look at it: Semantic

crossview localization and mapping with lidar. IEEE

Robotics and Automation Letters, 6:2397–2404.

Pan, Y., Zhong, X., Wiesmann, L., Posewsky, T., Behley,

J., and Stachniss, C. (2024). Pin-slam: Lidar slam

using a point-based implicit neural representation for

achieving global map consistency. IEEE Transactions

on Robotics, 40:4045–4064.

Razlaw, J., Droeschel, D., Holz, D., and Behnke, S. (2015).

Evaluation of registration methods for sparse 3d laser

scans. In 2015 European Conference on Mobile

Robots (ECMR), pages 1–7. IEEE.

Suger, B. and Burgard, W. (2017). Global outer-urban

navigation with openstreetmap. In 2017 IEEE In-

ternational Conference on Robotics and Automation

(ICRA), pages 1417–1422. IEEE.

Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L.,

Behley, J., and Stachniss, C. (2023). Kiss-icp: In de-

fense of point-to-point icp – simple, accurate, and ro-

bust registration if done the right way. IEEE Robotics

and Automation Letters, 8:1029–1036.

Xia, Z., Shi, Y., Li, H., and Kooij, J. F. P. (2024). Adapting

Fine-Grained Cross-View Localization to Areas With-

out Fine Ground Truth, pages 397–415.

Xu, W., Cai, Y., He, D., Lin, J., and Zhang, F. (2022). Fast-

lio2: Fast direct lidar-inertial odometry. IEEE Trans-

actions on Robotics, 38:2053–2073.

Yan, F., Vysotska, O., and Stachniss, C. (2019). Global lo-

calization on openstreetmap using 4-bit semantic de-

scriptors. In 2019 European Conference on Mobile

Robots (ECMR), pages 1–7. IEEE.

Yang, S., Jiang, R., Wang, H., and Ge, S. S. (2017).

Road constrained monocular visual localization using

gaussian-gaussian cloud model. IEEE Transactions

on Intelligent Transportation Systems, 18:3449–3456.

Yifan, D., Zhang, X., Li, Y., You, G., Chu, X., Ji, J., and

Zhang, Y. (2024). Cellmap: Enhancing lidar slam

through elastic and lightweight spherical map repre-

sentation.

Zheng, X. and Zhu, J. (2024a). Traj-lio: A resilient multi-

lidar multi-imu state estimator through sparse gaus-

sian process.

Zheng, X. and Zhu, J. (2024b). Traj-lo: In defense of

lidar-only odometry using an effective continuous-

time trajectory. IEEE Robotics and Automation Let-

ters, 9:1961–1968.

VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems

188