Semantic Segmentation and Effect Optimization of 3D Point Cloud Based on 2D Semantic Segmentation and Clustering for Construction Machinery Unstructured Environment.

The operational environment of construction machinery is predominantly unstructured, characterized by rapid changes, high complexity, and irregularly distributed objects. This poses significant challenges for 3D semantic perception, particularly due to the high cost of acquiring point cloud semantic labels. To address this, a novel 3D semantic perception scheme is proposed for such unstructured environments. This scheme integrates image semantic segmentation results with point cloud clustering via perspective projection. The projection parameters are refined using Particle Swarm Optimization (PSO), and the semantic consistency of the fused results is further enhanced by a Kd-tree-based radius nearest neighbor (RNN) matching algorithm. Consequently, a weakly supervised framework is established that achieves accurate 3D semantic understanding using only 2D image labels, eliminating the need for annotated 3D point clouds. The feasibility and effectiveness of the scheme are validated through a dedicated unstructured scene dataset and real-world testing. Results demonstrate its capability to effectively perceive 3D semantic information and reconstruct target contours, achieving a mean Pixel Accuracy (mPA) of 84.72% and a mean Intersection over Union (mIoU) of 75.85%.