Abstract
As large-scale laser 3D point clouds data contains massive and complex data, it faces great challenges in the automatic intelligent processing and classification of large-scale 3D point clouds. Aiming at the problem that 3D point clouds in complex scenes are self-occluded or occluded, which could affect the object classification accuracy, we propose a multi-dimensional feature optimization combination classification method named MFOC-CliqueNet based on CliqueNet for large-scale laser point clouds. The optimal combination matrix of multi-dimensional features is constructed by extracting the three-dimensional features and multi-directional two-dimensional features of three-dimensional point cloud. This is the first time that multi-dimensional optimal combination features are introduced into cyclic convolutional networks CliqueNet. The experimental results show that the MFOC-CliqueNet framework can realize the latest level with fewer parameters. The classification accuracy of our method is 98.9% on the Large-Scale Scene Point Cloud Oakland dataset, which is better than other classification algorithms mentioned in this paper.
Introduction
Classifying 3D point clouds in large-scale outdoor environments, such as urban streets, is a significant challenge due to data complexity and occlusion. Objects like trees, poles, and buildings often block one another, leading to incomplete data from certain perspectives. Existing classification methods typically either process 3D data directly—which can be computationally intensive and miss local geometric nuances—or project data onto a single 2D plane, losing vital structural information.
MFOC-CliqueNet addresses these issues by combining 3D spatial features with multi-angle 2D projections into an "Optimal Combination Matrix," which is then processed by a parameter-efficient cyclic convolutional network.
Methodology
The MFOC-CliqueNet framework consists of two primary stages: Multidimensional Feature Optimal Combination (MFOC) and the CliqueNet classifier.
Multidimensional Feature Optimal Combination (MFOC)
The core innovation lies in how features are extracted and combined:
- 3D Feature Extraction: Using a KD-Tree and Shannon entropy to determine the optimal neighborhood for each point, the system calculates descriptors such as local density, verticality, and linearity.
- 2D Multi-angle Projections: To mitigate the effects of occlusion, 3D points are projected onto three orthogonal planes: XOY (H-Plane), YOZ (w-Plane), and XOZ (v-Plane). This multi-view approach captures structural information that might be hidden in a single 3D view.
- Feature Integration: The researchers determined that a 9:1 ratio (favoring 3D features) and a specific "DCB" arrangement of projection data yielded the highest accuracy.
CliqueNet Classifier
Unlike standard feed-forward networks, CliqueNet utilizes cyclic feedback where the output of later layers is fed back into earlier ones. This allows the network to iteratively refine its feature understanding, achieving high accuracy with significantly fewer parameters than traditional deep learning architectures.
Experimental Results
The model was evaluated using the Oakland 3D Point Cloud Dataset, which contains urban categories such as Wire, Pole/Trunk, Facade, Ground, and Vegetation.
- Overall Accuracy: MFOC-CliqueNet achieved a state-of-the-art 98.9% accuracy.
- Comparative Performance: This result outperformed established methods like M3N (91.66%) and CRFoptN (95.5%).
- Class-Specific Strengths: The model was exceptionally strong at identifying Ground (99.5%) and Facade (94.6%) points.
Conclusion and Future Work
MFOC-CliqueNet demonstrates that 2D projections serve as a vital "safety net" for 3D features, providing context when spatial data is occluded. While the model is highly efficient and accurate, it still faces challenges in distinguishing between geometrically similar thin objects like poles and wires. Future research will focus on refining feature extraction for these specific categories to improve robustness in dense urban scenes.
Summary updated