Integrative and Multi-scale Deep Learning for 3D Point Cloud Transmission Corridor Scene Segmentation: Noise Filtering, Attention-Fused Feature Integration, and Panoptic Network

LiDAR technology plays a crucial role in mapping natural and built environments across various civic and military applications. It enables the acquisition of high-density 3D point clouds with pulse repetition frequencies ranging from 100Hz to 2MHz. However, the increased overlap with atmospheric points has posed challenges in noise filtering and 3D point cloud quality. This dissertation proposes the Noise Seeking Attention Network (NSANet), a novel solution integrating psychological theories of feature integration and attention engagement. NSANet achieves a 4.10% increase in F1-Score and a 7.30% improvement in recall by employing multiscale context, global physical priors, and local spatial attention for noise filtering, surpassing previous techniques.

The study explores the relationship between vegetation height and plantation guidelines, identifying spatial layout consistency in utility layouts and transmission objects. This insight drives the development of three semantic analysis approaches: Semantic Utility Network (SUNet), Fusion-Semantic Utility Network (Fusion-SUNet), and Panoptic-Semantic Utility Network (Pan-SUNet). Encouraged by the performance improvements of SUNet and Fusion-SUNet, Pan-SUNet achieves outstanding results, boasting a 94% F1-Score for pylons, 99% for ground, vegetation, and powerlines, and demonstrating high precision in 3D object detection.

Experiments conducted on Teledyne Optech's Galaxy T1000 dataset, which features diverse voltage transmission lines, validate the effectiveness of Pan-SUNet, particularly when combined with the RandLA baseline. Significant improvements are observed, including an increase from 80% to 85% in F1-Score for pylons, 98% to 99% for ground, 93% to 97% for powerline, 75% to 78.3% for other objects, 86% to 88% for buildings, while maintaining 98.2% for high vegetation and 93% for medium vegetation.

The key contribution of this research is a significant advancement beyond basic object classification, as it not only identifies the class of an object but also distinguishes between different instances of the same class. This instance segmentation is critical for utility network modelling and simulation.

The research emphasizes the importance of external cues such as contextual reasoning, spatial cognition, and physical priors for multiscale fusion in scene understanding systems. Moreover, the study's adaptability to integrate proposed contributions into existing networks enhances their overall performance, making them network-agnostic.