ibrahim80876/Change-Detection-LiDAR-Dataset

Subiaco LiDAR Change Detection Pairs (2023 to 2025)

Overview

This dataset provides 20 paired 3D point-cloud areas for urban change detection in Subiaco, an inner suburb of Perth, Western Australia. Each pair shows the same city block captured by mobile LiDAR in two surveys:

Survey	Date	Sensor	Source maps	Points
2023	15 July 2023	Ouster OS1-64 (64 beams)	5 SLAM maps	533,334,506
2025	26 February 2025	Ouster OS1-128 (128 beams)	20 SLAM loops	1,018,867,380

Within each pair, the two clouds are levelled, centred and co-registered so that unchanged surfaces (roads, walls, kerbs) coincide closely. The remaining differences represent change between the surveys, such as new or demolished buildings, construction, tree growth or removal, street furniture and parked vehicles.

The full source maps are published separately on IEEE DataPort (see Source datasets). This repository contains the derived pairs, the scripts used to produce them, and preview figures.

Repository structure

Change-Detection-LiDAR-Dataset/
├── data/
│   ├── Pair1/
│   │   ├── <2023 source map name> - Cloud.ply   # original 2023 crop, source-map frame
│   │   ├── <2025 source loop name> - Cloud.ply  # original 2025 crop, loop frame
│   │   ├── 2023.ply                             # levelled, centred, co-registered
│   │   ├── 2025.ply                             # levelled, centred, co-registered
│   │   └── pair_info.json                       # Pair9 to Pair20 only
│   ├── Pair2/
│   └── ... Pair20/
├── scripts/      # data preparation pipeline (Python)
├── figures/      # per-pair height maps and height-difference previews
└── README.md

Original crops are sections cut from the reconstructed SLAM point-cloud maps, not raw sensor recordings. They remain in the coordinate frame of their source map or loop.

Aligned files (2023.ply, 2025.ply) are the ready-to-use pair. Both are rotated to horizontal, centred on the pair area, and co-registered, with the 2025 ground at z = 0.

Pair numbers are area identifiers. They do not correspond to survey loop numbers.

File format

All point clouds are ASCII PLY files with the fields x y z r g b. Only the coordinates carry information; the RGB fields are zero, as in the source maps. There is no intensity or return information. Units are metres.

pair_info.json (Pair9 to Pair20) records every parameter and metric for the pair: crop boxes in both frames, transforms, tilt angles, ICP results and alignment residuals.

Pair summary

Pair	2023 source map	2025 loop (capture time)	Crop size (m)	2023 points	2025 points	Method
Pair1	20230715_1439	Loop3 (18:52)	86 x 83	1,297,020	4,138,268	Manual
Pair2	20230715_1450	Loop4 (18:55)	84 x 85	1,482,712	4,356,884	Manual
Pair3	20230715_1439	Loop3 (18:52)	33 x 32	372,836	810,816	Manual
Pair4	20230715_1450	Loop4 (18:55)	57 x 57	611,251	1,345,516	Manual
Pair5	20230715_1450	Loop6 (19:02)	88 x 87	1,815,359	2,880,173	Manual
Pair6	20230715_1514	Loop17 (19:40)	96 x 97	1,597,817	2,502,520	Manual
Pair7	20230715_1459	Loop17 (19:40)	74 x 74	1,231,858	1,691,443	Manual
Pair8	20230715_1439	Loop1 (18:45)	53 x 55	802,711	1,199,665	Manual
Pair9	20230715_1459	Loop7 (19:22)	80 x 80	2,410,808	3,691,667	Automated
Pair10	20230715_1450	Loop3 (18:52)	80 x 80	3,444,087	6,549,773	Automated
Pair11	20230715_1514	Loop21 (unknown)	80 x 80	1,857,735	2,060,814	Automated
Pair12	20230715_1459	Loop7 (19:22)	80 x 80	2,179,254	3,795,109	Automated
Pair13	20230715_1514	Loop21 (unknown)	80 x 80	2,273,353	2,115,457	Automated
Pair14	20230715_1459	Loop8 (19:20)	80 x 80	1,651,922	2,998,040	Automated
Pair15	20230715_1459	Loop7 (19:22)	80 x 80	1,463,816	2,542,975	Automated
Pair16	20230715_1450	Loop4 (18:55)	80 x 80	1,965,612	4,896,852	Automated
Pair17	20230715_1514	Loop21 (unknown)	80 x 80	1,772,992	1,929,153	Automated
Pair18	20230715_1450	Loop6 (19:02)	80 x 80	1,173,944	2,659,985	Automated
Pair19	20230715_1514	Loop17 (19:40)	80 x 80	1,646,048	2,287,563	Automated
Pair20	20230715_1450	Loop5 (18:59)	80 x 80	1,439,820	4,176,673	Automated

For Pair9 to Pair20, the 2025 point counts are after clipping to the aligned 2023 footprint.

How the pairs were made

Pair1 to Pair8 (manual). Areas were cropped in CloudCompare from an earlier co-registered export, then levelled and re-centred in Rhino 6 using a rigid transform per cloud. Their positions in the current source maps were recovered afterwards by relief correlation.

Pair9 to Pair20 (automated). The source maps and loops each use their own local SLAM frame and are tilted by 2 to 12 degrees, so each pair required its own registration. The pipeline:

1. Summarises every source cloud on a 2 m grid (point count, minimum and maximum height).
2. Estimates the 2D rigid transform (yaw, x, y) between each 2023 map and 2025 loop by masked normalised cross-correlation (NCC) of height-range images over all yaw angles.
3. Selects 80 m x 80 m candidate areas with at least 60% joint coverage and 300,000 points per epoch, spaced at least 130 m apart and away from existing pairs. A flatness test rejects streets affected by SLAM "ghost layers" (duplicate surfaces displaced 5 to 10 m vertically).
4. Verifies each candidate by searching for its 2023 height pattern across the whole 2025 loop, accepting only unambiguous matches (NCC of at least 0.70, a margin of at least 0.15 over the second-best peak, and within 150 m of the predicted position).
5. Crops both epochs from the source files.
6. Levels each cloud using a RANSAC ground plane, centres it, applies a coarse shift by DSM cross-correlation, and refines with trimmed point-to-point ICP.

As a check, the automated pipeline was run on the original crops of Pair1, Pair6 and Pair7 and reproduced the manual alignment to within 0.10 to 0.15 m.

Alignment quality

Alignment is measured as the median nearest-neighbour distance between the aligned 2023 and 2025 clouds:

• Manual pairs (Pair1 to Pair8): 0.08 to 0.15 m, except Pair6 at about 0.5 m.
• Automated pairs (Pair9 to Pair20): 0.15 to 0.30 m (2023 to 2025 direction).

These distances include genuine change, vegetation, vehicles and occlusion, so they are an upper bound on registration error for static surfaces.

Data preparation scripts

The scripts/ folder contains the full pipeline:

Script	Purpose
footprint.py	Summarises source clouds on a 2 m grid
inventory.py	Lists source files, point counts, sizes and coverage
analyse_pairs.py	Checks the original crops and their alignment transforms
locate_existing.py	Finds existing crops within the source maps
pairwise_frames.py	Estimates the relative position and rotation of maps from the two years
select_verify.py	Selects and checks additional matching areas
crop.py	Extracts selected areas from the source clouds
align.py	Levels, centres and aligns each pair
make_pairs.py	Runs cropping and alignment and records pair metadata
make_figures.py	Creates paired height maps and height-difference previews
overview2.py	Shows where the selected areas lie within the 2023 source maps

Usage

Download a single pair and load it with Open3D:

from huggingface_hub import snapshot_download
import open3d as o3d

path = snapshot_download(
    repo_id="ibrahim80876/Change-Detection-LiDAR-Dataset",
    repo_type="dataset",
    allow_patterns=["data/Pair9/2023.ply", "data/Pair9/2025.ply"],
)

pc_2023 = o3d.io.read_point_cloud(f"{path}/data/Pair9/2023.ply")
pc_2025 = o3d.io.read_point_cloud(f"{path}/data/Pair9/2025.ply")
print(pc_2023, pc_2025)

Download everything with the CLI:

hf download ibrahim80876/Change-Detection-LiDAR-Dataset --repo-type=dataset --local-dir ./subiaco_cd

Source datasets

The full source maps are available on IEEE DataPort:

• 2023 survey: 2023 Subiaco WA 3D HD LiDAR Point Cloud Maps Dataset
• 2025 survey: 2025 Subiaco WA 3D HD LiDAR GNSS Point Cloud Maps Dataset

Citation

This dataset accompanies the following paper, currently under review at Scientific Reports. If you use the dataset, please cite the paper and the source datasets.

Hezam Albaqami, Haitian Wang, Xinyu Wang, Muhammad Ibrahim, Zainy M. Malakan, Abdullah M. Algamdi, Mohammed H. Alghamdi and Ajmal Mian. "LiDAR-based 3D Change Detection at City Scale." Scientific Reports (under review).

@article{albaqami2026lidarcd,
  author  = {Albaqami, Hezam and Wang, Haitian and Wang, Xinyu and Ibrahim, Muhammad and Malakan, Zainy M. and Algamdi, Abdullah M. and Alghamdi, Mohammed H. and Mian, Ajmal},
  title   = {LiDAR-based 3D Change Detection at City Scale},
  journal = {Scientific Reports},
  year    = {2026},
  note    = {Under review}
}

Contact

Dr. Muhammad Ibrahim. Please use the Community tab for questions or issues.