A Time- and Memory-Efficient Method for Axis Alignment of
Very Large Point Clouds using Orthogonality in Buildings
Longyong WU∗ , Sou-Han CHEN , Meng SUN and Fan XUE
Department of Real Estate and Construction, The University of Hong Kong, Hong Kong (China SAR)
wulongyong@connect.hku.hk, hankchen@connect.hku.hk, sunmhku01@connect.hku.hk, xuef@hku.hk
modeling, 3D coordination, registration of 2D computeraided design (CAD) drawings [6], scan-to-BIM [7], scanvs-BIM for change detection [8], and construction progress
monitoring [9]. In computer vision, many deep learning
models for point cloud processing have been proposed and
trained well on coordinate-aligned datasets [10], and some
robots for building operation and maintenance also relied
on axis alignment.
However, existing tools for aligning LiDAR cloud coordinates with buildings and construction often encounter
challenges. Building floor areas and construction sites can
span multiple thousand square meters, and modern LiDAR
scanners produce 10,000–400,000 points per square meter; these factors result in LiDAR clouds that can reach
up to billions of points and tens or even hundreds of gigabytes (GB). Existing 3D point cloud processing tools, such
as Open3D [11] and pypcd4, always load the full cloud
data into computer memory in a single batch. This onebatch processing of very large LiDAR clouds is memoryintensive and sometimes exceeds the capacity of computers, which are equipped with no more than 32 GB
Random-access memory (RAM) in most cases. When
memory demand exceeds system capacity, the operating
system may terminate the process, causing the computation to fail and limiting the scalability of LiDAR data
analysis.
To mitigate the problem of exceeding system memory
capacity in axis alignment, this study presents a chunkbased, orthogonality-enabled parallel computing method.
The method aims to achieve relatively efficient axis alignment for very large point clouds. The method leverages the
random-access characteristic of LiDAR’s LASer (.las) format to divide an input LiDAR point cloud into partitioned
chunks, then processes them in parallel with controlled
concurrency. To ensure balanced orthogonality detection,
the point distances in each chunk are normalized (downsampled) to 2 cm, and point normals are estimated from the
normalized points. Finally, the method automatically estimates and optimizes the dominant orthogonal structural
directions of LiDAR point clouds, in order to rotate the
point cloud and align the coordinates. Overall, this parallel computing utilizing orthogonality in buildings can
achieve better time and memory efficiency on daily per-

Abstract In construction and operational stages of buildings and in
computer vision and construction robotics, accurate alignment of LiDAR point clouds to a building’s 3D coordinate
axes is an essential preprocessing step. However, existing
point cloud processing tools handle a point cloud in a single batch, which can lead to high memory usage or even
failures for very large clouds. This study presents a chunkbased, orthogonality-enabled method for aligning very large
building LiDAR point clouds to the building’s coordinate
in parallel computing. Our method leverages the randomaccess characteristic of LiDAR’s LASer (.las) format to divide a cloud into multiple chunks, then processes each chunk,
with the number of concurrently running tasks being limited. Subsequently, the global orientation is estimated by
extracting the dominant directions of the building’s interior
and exterior orthogonal structures to rotate and align the
input cloud. Experiments on a 22 GB real-world LiDAR
point cloud demonstrated the effectiveness of our proposed
method: It reduced peak memory consumption from exceeding 60.5 GB to 4.1 GB and achieved a 3.85-fold speedup
over the sequential method. (The source code is shared at:
https://github.com/chunibyo-wly/laspy-parallel)

Keywords Building and Construction; LiDAR; Point cloud processing; Orthogonal structure; Parallel computing

1

Introduction

Light Detection and Ranging (LiDAR) point clouds
have become increasingly essential data sources in construction [1], surveying [2], computer vision [3], and
robotics [4], thanks to LiDAR’s high precision in measuring the 3D built environment. Yet, a LiDAR point cloud’s
local coordinate axes are usually inconsistent with the target building’s or project’s coordinate axes in 2D drawings,
footprint on the map, or 3D building information modeling
(BIM).
Axis alignment refers to aligning a point cloud to the
project’s local coordinate axes [5]. Axis alignment is an
important step to enable many construction digitalization
and automation applications, including existing condition
1

1626

sonal computers.

3

Our Method for Processing Very Large
Point Clouds

3.1

LiDAR Scanning Principles and the LASer Format

The contribution of this paper is twofold:

1. A novel chunk-based, orthogonality-enabled parallel
LiDAR has been widely used in construction, robotics
computing method for axis alignment of very large
and
mapping. It actively emits laser pulses and measures
LiDAR point clouds, and
the time delay of their reflections from surfaces to estimate distances. Compared with passive sensors such as
2. Preliminary evidence on the boosted time and mem- cameras, LiDAR is less affected by lighting conditions
ory efficiencies of the present method from experi- and can provide depth information even in poor lighting. After obtaining raw distance measurements, LiDAR
ments on real-world very large data.
point clouds can be combined with simultaneous localization and mapping algorithms to automatically estimate the
scanner’s position and orientation in unknown environments, reconstructing 3D environment with rich textures
2 Related Works
and high-precision geometric details [18]. This capabilLiDAR has been widely applied in the construction ity makes LiDAR a powerful tool for change detection,
industry [12, 13] for various applications, ranging from building modeling, and progress monitoring [13].
To efficiently manage various and massive LiDAR
semantic enrichment of BIM models to derivative-free
point
clouds, The American Society for Photogrammeoptimization-based (DFO) 3D reconstruction of architectry
and
Remote Sensing defined an open format, LiDAR
tural structures. For instance, Xue et al. [14] is the first to
(with
“.las”
file extension), for the exchange and stortransform architectural symmetry detection from LiDAR
age
of
LiDAR
data. This format typically consists of
point clouds as a nonlinear optimization problem and profour
components:
Public Header Block, Variable-Length
posed a DFO approach to solve it. Furthermore, Xue et al.
Records,
Point
Data
Records, and Extended Variable[15] extended the DFO method to register a small-scale inLength
Records.
The
header
block contains the metadata
door LiDAR point clouds with BIM components, enabling
indicating
the
byte
offset
to
the
beginning of point data and
the reconstruction of detailed indoor scenes.
the byte size of each point, which can be used to access
Although existing studies have demonstrated LiDAR
each point efficiently [19]. This benefits in parallel and
point clouds’ applications in 3D modeling and semandistributed algorithms for very large point cloud processtic enrichment, most have been limited to small-scale or
ing.
exterior-level point clouds [14, 16], due to computational
and memory limits. In practice, LiDAR point clouds collected from construction sites often contain millions of 3.2 Chunk-based Parallel Processing
interior points, posing significant challenges for preproExisting point cloud processing tools, such as Open3D
cessing using existing methods. Nevertheless, these inteand pypcd4, always load a LiDAR point cloud in a sinrior point clouds are valuable for analyzing architectural
gle batch. However, this approach is limited in handling
symmetries and skeletons.
very large LiDAR data, as loading the whole file in a sinOrthogonality, a common characteristic of real-world gle batch can easily cause out-of-memory failures. For
architectural forms and designs, offers researchers an op- instance, Open3D, a well-known point cloud processing liportunity to understand and reconstruct urban point clouds brary, stores point coordinates, colors, and normals using
[17]. For instance, Wu et al. [12] applied orthogonal- std::vector<Eigen::Vector3d> in memory, which is
ity and symmetry constraints to efficiently reconstruct 3D memory-intensive. Loading 789-million points into memmodels from facade point clouds, and later extended this ory at once requires approximately 53 GB before any furframework to support morphological analysis. This inher- ther processing. Such memory demand exceeds the capacent architectural characteristic enables the development of ity of computers and even some workstations. Therefore,
chunk-based, orthogonality-enabled methods for the axis it is essential to develop a memory-efficient method to
alignment of very large LiDAR point clouds.
deal with massive point clouds while maintaining time
In summary, it is of interest to design a chunk-based, efficiency.
Leveraging the structure of the LASer format, a point
orthogonality-enabled method for axis alignment of very
large LiDAR point clouds, for LiDAR and construction cloud can be randomly accessed using the following fordigitalization practitioners who handle very large indoor mula:
or outdoor scenes in a time- and memory-efficient manner.
Offset = Offsetheader + 𝑖 × TypeSize
(1)

1627

where 𝑖 denotes the index of the specific point to be accessed directly, and TypeSize denotes the number of bytes
used to store each point (coordinates and attributes) as defined in the LASer header. This random-access property
enables data access without sequential reading, allowing
the LiDAR point clouds to be divided into smaller chunks
to trade off between memory efficiency and processing
time.
However, processing these chunks in parallel straightforwardly introduces a new challenge: each thread requires dedicated memory allocation before execution. If
all chunks are loaded and processed simultaneously, it will
still raise an out-of-memory error. This is because earlier
tasks may not finish in time to release their allocated memory before later tasks begin to require additional memory.
To address this issue, our chunk-based parallel processing
approach divides tasks into multiple batches (Fig. 1) and
limits the number of concurrent processes. The next batch
is launched after the slowest task in the previous batch
finishes.

rotation estimation only around the 𝑧-axis. Specifically, by
projecting each surface normal n𝑖 = (𝑛𝑥𝑖, 𝑛 𝑦𝑖 , 𝑛 𝑧𝑖 ) onto
the 𝑥𝑦-plane, we can describe its horizontal orientation
using a single rotation angle:
𝜃 𝑖 = arctan

𝑛 𝑦𝑖 180
×
.
𝑛 𝑥𝑖
𝜋

(4)

The whole angles range from −180◦ to 180◦ . We quantized the float64 values to uint16 integers by scaling
with a factor of 100 to reduce memory usage. This is because 216 = 65, 536 > 360 × 100, all the possible angles
can be represented with two-decimal precision. Therefore, three floating-point values representing normals can
be compactly stored as integers without exceeding memory limits.
After chunk-based parallel processing, all rotation angles are collected in the range [−180◦ , 180◦ ). Since structural components such as walls, windows, and facades are
often designed to fit orthogonal layouts (Fig. 3), their surface normals are distributed around 4 principal directions
separated by 90◦ in the 𝑥𝑦-plane. Leveraging this centralsymmetric property, the angle distribution can be folded
into the reduced range [0◦ , 90◦ ) to aggregate geometrically
equivalent orientations, as illustrated in Fig. 2.
𝜃 in [−180◦ , 180◦ )

If 𝜃 < 0◦ , then
𝜃 = 𝜃 + 180◦

Final range [0◦ , 90◦ )

If 𝜃 > 90◦ , then
𝜃 = 𝜃 − 90◦

Figure 1. Illustration of chunk-based parallel processing

3.3

Axis Alignment of LiDAR point clouds

This section details the three steps of the axis alignment
method that can avoid massive memory demands.

Figure 2. Range reduction from [−180◦ , 180◦ ) to
[0◦ , 90◦ )

Normal estimation (Parallel computing). In the first
step, we compute the normal vectors for each point. Given
a point 𝑝 𝑖 , the set of neighboring points within a radius 𝑟
are defined as 𝑃 = [ 𝑝 1 , 𝑝 2 , ..., 𝑝 𝑖 , ...𝑝 𝑘 ]. The covariance
matrix of point 𝑝 𝑖 is estimated as
𝐶 ( 𝑝𝑖 ) =

1 𝑇
𝑃𝑃
𝑘

(2)

The corresponding normal vector 𝑛𝑖 is then obtained using
Singular Value Decomposition (SVD):
𝑛𝑖 = SVD(𝐶)

(3)

Figure 3. Illustration of the orthogonal structural
components (e.g., walls) in LiDAR point clouds

Rotation Estimation. Since the gravity direction is usually accurately provided by LiDAR scanners, the 3D rota- Peak Direction Estimation In the final step, we extract
tion in axis alignment problem can be simplified to a 2D the peak angle with the highest frequency in the range

1628

[0◦ , 90◦ ). This angle denotes the dominant rotation angle
between the current point cloud orientation and the global
𝑥𝑦-axis. By rotating the point cloud around the 𝑧-axis by
this estimated angle, the orthogonal building component
can be parallel to the 𝑥𝑦-axis. Note that a finer resolution
in the bins of degrees can lead to more accurate peak
direction estimation.

estimation, and (3) axis alignment. We focus on these
stages because they account for most of the memory demand throughout the point cloud preprocessing pipeline
and often serve as the foundation for subsequent point
cloud analyses. Therefore, evaluating these three stages
can provide a comprehensive assessment of both memory
efficiency and computational efficiency in very large point
cloud processing.
4.2

Results

We compared the proposed method with two baseline
implementations, Open3D2 and pypcd43 , as well as a sequential chunks version of our method. As shown in Table 1, our proposed method significantly improves memory
usage while maintaining computational efficiency.
Table 1. Comparison of methods for aligning the
789-million point cloud
Method

Experiments

4.1

Experimental Setup

To evaluate the proposed chunk-based parallel processing method for massive LiDAR point clouds, we conducted experiments on a real-world LiDAR scan dataset.
The case study involves a real-world LiDAR point cloud
dataset collected from the Hang Tai Road, Ma On Shan
Area 86B Phase 2 construction project, a public housing
development under the Hong Kong Housing Authority1 .
The scanned area covers approximately 3,392 m2 . The
data were acquired using a NavVis VLX 3 system, which
provides 5 mm accuracy and a capture rate of 2 × 640, 000
points/s. The resulting LAS file is 22 GB in size and contains 789,448,875 points, with an average point density of
approximately 232,738 points/m2 .
All experiments were evaluated on an HP Z2 workstation equipped with an Intel Core i7-13700K CPU, 64 GB
RAM with 16 GB swap memory, running under the Windows Subsystem for Linux (WSL) environment.
Our evaluation focuses on the pipeline from reading
the point cloud from disk to rotating it into alignment
with the global coordinate axes. Specifically, there are
three consecutive stages: (1) data loading, (2) normal
1 Approved Planning Brief from the Hong Kong Housing Authority

Peak RAM ↓

Open3D Baseline
N.A.†
>64 GB†
pypcd4 Baseline
27.0 mins
60.5 GB
Sequential Chunks 98.6 mins
4.9 GB
Ours (this paper) 25.6 mins
4.1 GB
↓: Smaller is better; †: Failed at 23 mins.

Figure 4. Degrees distribution of LiDAR point
clouds normals in 𝑥𝑦-plane

4

Runtime ↓

Specifically, as illustrated in Fig. 5, Open3D takes
around 30 minutes to load the point cloud into memory,
occupying approximately 53.4 GB the system’s available
memory, and continues to allocate a large amount of memory, causing the process to be terminated by the system
process manager. Compared with Open3D, which uses
the float64 format, pypcd4 stores point coordinates in
float32 format in RAM for reducing both the memory
demand (60.5 GB) and the time (27 mins) spent on memory allocation during evaluation.
The synchronization chunk processing method sequentially processes the chunk’s point cloud, significantly reducing peak memory usage (Fig. 5). However, the total
runtime is around 1 h 38 mins. In contrast, our proposed
parallel chunk-based method achieves a trade-off between
memory and time efficiency (Fig. 5). The whole process
finishes within 26 minutes, which is 3.85-fold speedup
compared with the synchronization chunk-based method.
And the peak memory usage is also below 5 GB, considerably lower than the system’s limits.
In addition to total runtime and peak RAM usage, Fig. 6
breaks the runtime into two stages, (1) data loading, (2)
normal estimation and axis alignment. The results show
2 https://github.com/isl-org/Open3D
3 https://github.com/MapIV/pypcd4

1629

Figure 5. Memory usage comparison of Open3D (single batch), pypcd4 (single batch), sequential chunk processing, and our proposed method

that most of the runtime is spent on data loading, indicating that I/O is the major bottleneck in this pipeline.
Compared with the sequential chunk way, our proposed
method significantly reduces the time spent on point cloud
loading from disk, achieving a 4.98-fold speedup. This
demonstrates that our proposed method efficiently avoids
memory accumulation while maintaining time efficiency.

Axis Alignment Result After parallel processing, the
point cloud was successfully aligned to the global 3D axis,
as illustrated in Fig. 7. The red, green, and blue arrows
represent the 𝑥𝑦𝑧-axis, respectively. The result demonstrates that our estimated dominant rotation angle enables
efficient alignment of very large LiDAR point clouds to
global coordinates.
Fig. 8 shows the registration result of LiDAR point
clouds with the 2D drawings. As the point clouds have already been globally axis-aligned, the registration process
only requires solving 2 degrees of freedom (translation
along the 𝑥𝑦-axes), significantly reducing human effort
compared with full 3D registration.

1630

Figure 6. Runtime break down comparison of
pypcd4 (single batch), sequential chunk processing, and our proposed method

Figure 7. Axis alignment result of the LiDAR point clouds.

Figure 8. Registration result of 2D drawings with LiDAR point clouds. For better visualization, a sector along
the 𝑧-axis is selected.

1631

5

Discussion and Limitations

6

Conclusion

and Stanford polygon “.PLY”, (2) analyze the influence
of chunk size and concurrency on the memory and CPU
The experimental results demonstrated that the pro- utilization, and (3) evaluate the method on a wider range
posed chunk-based, orthogonality-enabled parallel com- of very large scanning datasets.
puting method effectively estimates the global orientation
of very large LiDAR point clouds and achieves both time
and memory efficiencies. Compared with the single-batch Acknowledgment
loading approach in pypcd4, our method significantly reThe data used in this study were collected and produces peak memory usage from over 60.5 GB to approxi- vided by Geosys Hong Kong Limited. The work demately 4.1 GB while maintaining comparable data loading scribed in this study was supported by the Hong Kong
time. These findings demonstrate that loading a very large Research Grants Council (No. 17201325) and, in part, by
point cloud into memory at once is inefficient when mem- the Hong Kong Innovation and Technology Commission
ory is limited. In contrast, aggregating local geometric (No. ITP/004/23LP).
cues from chunks without full memory access to the full
point cloud can be a more efficient way to estimate global
References
orientation and preprocess.
However, several limitations still exist. First, the pro- [1] Mingyu Zhang, Yushu Yang, Shuai Han, Heng
posed method currently supports only the LASer file forLi, Dongliang Han, Xiaotong Yang, and Nan Guo.
mat, although “.PLY” and “.PCD” also have randomLidar-based framework for accurate positioning and
access properties. Second, the hyperparameters, such as
robust tracking of multiple construction workers.
the chunk size and the number of concurrent processes, are
Journal of Computing in Civil Engineering, 39(3):
set manually, leading to underutilization of RAM and CPU
04025027, 2025.
resources. Last but not least, the method was not tested
across diverse building and construction scenes. More im- [2] Paul Sestras, Gheorghe Badea, Ana Cornelia Badea,
Tudor Salagean, Sanda Ros, ca, Shuraik Kader, and
portantly, the assumption of dominant orthogonality may
Fabio Remondino. Land surveying with uav phonot fit for all buildings (e.g., curved facades and nontogrammetry and lidar for optimal building planning.
Manhattan layouts). In such cases, the degree distribution
Automation in Construction, 173:106092, 2025.
of surface normals may have multiple peaks, potentially
degrading the global orientation estimation and failing the
[3] Wenyao Liu, Jinhua Chen, Zemin Lyu, Rui Feng,
axis alignment. In addition, this study does not provide exTong Hu, and Lu Deng. Automatic tile position and
periments on how chunk size affects runtime and memory
orientation detection combining deep-learning and
usage, which should be further explored in future work.
rule-based computer vision algorithms. Automation
in Construction, 171:106001, 2025.
This study presents a chunk-based, orthogonalityenabled method for automatically aligning the global axes
of very large LiDAR point clouds. By leveraging the
random-access property of the LASer file and controlling parallel concurrency, our proposed method efficiently
achieves massive point cloud loading, normal estimation,
and global axis alignment on very large LiDAR point
clouds on computers. Experiments are conducted on a
22 GB real-world LASer file. The experimental results
demonstrated that our method drastically reduces memory
consumption and achieves a 3.85-fold speedup compared
with the sequential chunk-based approach. The parallel
computing mechanisms in this axis alignment method can
also benefit other applications of LiDAR point clouds, including learning-based point cloud analysis, point cloudto-CAD/BIM registration, change detection, and construction progress monitoring.
For future work, one can (1) extend the support for
point cloud exchange formats, such as pypcd4’s “.PCD”

1632

[4] Andrew Yarovoi and Yong Kwon Cho. Review of
simultaneous localization and mapping (slam) for
construction robotics applications. Automation in
Construction, 162:105344, 2024.
[5] Feng Li, Wenzhong Shi, Yunlin Tu, and Hua Zhang.
Automated methods for indoor point cloud preprocessing: Coordinate frame reorientation and building exterior removal. Journal of Building Engineering, 76:107270, 2023.
[6] Lin Bie, Shouan Pan, Siqi Li, Yining Zhao, and Yue
Gao. Graphi2p: Image-to-point cloud registration
with exploring pattern of correspondence via graph
learning. In Proceedings of the Computer Vision
and Pattern Recognition Conference, pages 22161–
22171, 2025.
[7] Ke You, Ke Chen, and Fan Xue. Automated scan-tobim for construction digital transformation: Conceptual framework, processing methods and best-

practice guidelines. Automation in Construction,
182:106755, 2026.

efficient convex decomposition method of 3d building models for urban morphological analytics. International Journal of Geographical Information Science, pages 1–23, 2025.

[8] Iris De Gélis, Thomas Corpetti, and Sébastien
Lefèvre. Change detection needs change information: Improving deep 3-d point cloud change detec- [17] Xiaolu Zeng, Yang Hu, Xiaopeng Yang, Zixiang Yin,
and Shichao Zhong. A multi-source fusion system
tion. IEEE Transactions on Geoscience and Remote
for through-wall radar compensation using lidar and
Sensing, 62:1–10, 2024.
slam-based 3d reconstruction. IEEE Sensors Jour[9] Dong Liang, Longyong Wu, Meng Sun, Ruibo Hu,
nal, 2025.
Lingming Kong, Yipeng Pan, and Fan Xue. Transfer learning from building information model-based [18] Chunran Zheng, Wei Xu, Zuhao Zou, Tong Hua,
Chongjian Yuan, Dongjiao He, Bingyang Zhou,
synthetic data for three-dimensional module detecZheng Liu, Jiarong Lin, Fangcheng Zhu, et al. Fasttion in point clouds of modular-integrated construclivo2: Fast, direct lidar-inertial-visual odometry.
tion hoisting. Engineering Applications of Artificial
IEEE Transactions on Robotics, 2024.
Intelligence, 164:113243, 2026.
[10] Maxim Kolodiazhnyi, Anna Vorontsova, Anton [19] Andre Samberg. An implementation of the asprs las
standard. In ISPRS Workshop on Laser Scanning and
Konushin, and Danila Rukhovich. Oneformer3d:
SilviLaser, pages 363–372, 2007.
One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition, pages
20943–20953, 2024.
[11] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun.
Open3D: A modern library for 3D data processing.
arXiv:1801.09847, 2018.
[12] Yijie Wu, Fan Xue, Maosu Li, and Sou-Han Chen. A
novel building section skeleton for compact 3d reconstruction from point clouds: A study of high-density
urban scenes. ISPRS Journal of Photogrammetry
and Remote Sensing, 209:85–100, 2024.
[13] Dong Liang, Sou-Han Chen, Zhe Chen, Yijie Wu,
Louis YL Chu, and Fan Xue. 4d point cloud-based
spatial-temporal semantic registration for monitoring
mobile crane construction activities. Automation in
Construction, 165:105576, 2024.
[14] Fan Xue, Weisheng Lu, Christopher J Webster, and
Ke Chen. A derivative-free optimization-based approach for detecting architectural symmetries from
3d point clouds. ISPRS Journal of Photogrammetry
and Remote Sensing, 148:32–40, 2019.
[15] Fan Xue, Weisheng Lu, Ke Chen, and Anna Zetkulic.
From semantic segmentation to semantic registration: derivative-free optimization–based approach
for automatic generation of semantically rich as-built
building information models from 3d point clouds.
Journal of Computing in Civil Engineering, 33(4):
04019024, 2019.
[16] Yijie Wu, Fan Xue, Liangliang Nan, Longyong Wu,
Jantien Stoter, and Anthony GO Yeh. Morphcut: an

1633