Abstract:A 3D point cloud is an unstructured, sparse, and irregular dataset, typically collected by airborne LiDAR systems over a geological region. Laser pulses emitted from these systems reflect off objects both on and above the ground, resulting in a dataset containing the longitude, latitude, and elevation of each point, as well as information about the corresponding laser pulse strengths. A widely studied research problem, addressed in many previous works, is ground filtering, which involves partitioning the points into ground and non-ground subsets. This research introduces a novel task: detecting and identifying human-made objects amidst natural tree structures. This task is performed on the subset of non-ground points derived from the ground filtering stage. Marked Point Fields (MPFs) are used as models well-suited to these tasks. The proposed methodology consists of three stages: ground filtering, local information extraction (LIE), and clustering. In the ground filtering stage, a statistical method called One-Sided Regression (OSR) is introduced, addressing the limitations of prior ground filtering methods on uneven terrains. In the LIE stage, unsupervised learning methods are lacking. To mitigate this, a kernel-based method for the Hessian matrix of the MPF is developed. In the clustering stage, the Gaussian Mixture Model (GMM) is applied to the results of the LIE stage to partition the non-ground points into trees and human-made objects. The underlying assumption is that LiDAR points from trees exhibit a three-dimensional distribution, while those from human-made objects follow a two-dimensional distribution. The Hessian matrix of the MPF effectively captures this distinction. Experimental results demonstrate that the proposed ground filtering method outperforms previous techniques, and the LIE method successfully distinguishes between points representing trees and human-made objects.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to detect and locate human - made objects in 3D point - cloud data. Specifically, the author proposes an unsupervised machine - learning method, aiming to distinguish human - made objects in natural tree structures. The challenges of this problem are as follows: 1. **Characteristics of 3D Point - Cloud Data**: 3D point - cloud data is an unstructured, sparse and irregular data set, usually collected by airborne LiDAR systems. These data contain longitude, latitude, elevation of each point and laser pulse intensity information. 2. **Ground Filtering Problem**: It is necessary to separate ground points and non - ground points in the point - cloud data, which is the basis for many subsequent tasks. Traditional methods perform poorly on complex terrains. 3. **Lack of Labeled Data**: Due to the lack of labeled data, supervised learning methods cannot be used to deal with this problem. ### Method Overview To solve the above problems, the paper proposes a three - stage method: 1. **Ground Filtering**: - A new method called One - Sided Regression (OSR) is introduced to deal with the ground filtering problem on uneven terrains. - OSR assumes that ground points are located above the central plane, and non - ground points are regarded as outliers. The ground model parameters are estimated by minimizing positive residuals. 2. **Local Information Extraction (LIE)**: - The Hessian matrix of Marked Point Field (MPF) is estimated using the kernel method. - The tree points with three - dimensional distribution and the human - made object points with two - dimensional distribution are distinguished by analyzing the eigenvalues of the Hessian matrix. The specific formula is as follows: \[ H(p)=\begin{pmatrix} \frac{\partial^{2}\lambda(p)}{\partial x^{2}}&\frac{\partial^{2}\lambda(p)}{\partial x\partial y}&\frac{\partial^{2}\lambda(p)}{\partial x\partial z}\\ \frac{\partial^{2}\lambda(p)}{\partial y\partial x}&\frac{\partial^{2}\lambda(p)}{\partial y^{2}}&\frac{\partial^{2}\lambda(p)}{\partial y\partial z}\\ \frac{\partial^{2}\lambda(p)}{\partial z\partial x}&\frac{\partial^{2}\lambda(p)}{\partial z\partial y}&\frac{\partial^{2}\lambda(p)}{\partial z^{2}} \end{pmatrix} \] where \(p = (x, y, z)^{\top}\in D\) is an arbitrary point. 3. **Clustering**: - The Gaussian Mixture Model (GMM) is used to cluster the results of the LIE stage, and the non - ground points are divided into two categories: trees and human - made objects. ### Experimental Results The experimental results show that the proposed OSR method is superior to traditional methods in ground filtering, and the LIE method can successfully distinguish between trees and human - made objects. By combining these two methods, human - made objects can be effectively detected and located in complex 3D point - cloud data. ### Summary The main contribution of this research is to propose an unsupervised machine - learning framework that can accurately detect and locate human - made objects from 3D point - cloud data without labeled data. This method is not only applicable to natural environments such as forests, but also can be extended to fields such as urban planning and geological exploration.

Unsupervised Machine Learning for Detecting and Locating Human-Made Objects in 3D Point Cloud

Towards Open-Set Semantic Labeling in 3D Point Clouds : Analysis on the Unknown Class.

A Semantic Labeling Strategy to Reject Unknown Objects in Large Scale 3d Point Clouds

Surface Extraction and Boundary Detection Based on DBSCAN Clustering in 3D Point Clouds

DC3DCD: unsupervised learning for multiclass 3D point cloud change detection

Deep Unsupervised Learning for 3D ALS Point Cloud Change Detection

CFPC: The Curbed Fake Point Collector to Pseudo-LiDAR-Based 3D Object Detection for Autonomous Vehicles

MS23D: A 3D object detection method using multi-scale semantic feature points to construct 3D feature layer

Classification of airborne 3D point clouds regarding separation of vegetation in complex environments

Self-supervised 3D Human Mesh Recovery from Noisy Point Clouds

A Novel Method for 3D Object Detection in Open-Pit Mine Based on Hybrid Solid-State LiDAR Point Cloud

Application of Machine Learning to Detect Building Points in Photogrammetry-based Point Clouds

Towards A Weakly Supervised Framework for 3D Point Cloud Object Detection and Annotation

Automated Classification of Airborne Laser Scanning Point Clouds

Vision-Language Guidance for LiDAR-based Unsupervised 3D Object Detection

Roadside LiDAR Vehicle Detection and Tracking Using Range and Intensity Background Subtraction

3D Multi-object Detection and Tracking with Sparse Stationary LiDAR

Learning Occupancy for Monocular 3D Object Detection

Not All Points Are Equal: Learning Highly Efficient Point-based Detectors for 3D LiDAR Point Clouds

From Multi-View to Hollow-3D: Hallucinated Hollow-3D R-CNN for 3D Object Detection

Towards Unsupervised Object Detection From LiDAR Point Clouds