Instance-Level Semantic-Geometric Fusion for Cross-Domain Pole Detection in Railway LiDAR Point Clouds

Sammanfattning

Automatic mapping of catenary poles in LiDAR point clouds is increasingly important for railway infrastructure monitoring. However, existing pole detection methods often degrade under domain shift, limiting their generalization across datasets. This thesis investigates whether instance-level semantic-geometric late fusion can improve the robustness of pole detection in cross-domain railway LiDAR data. The proposed approach combines geometry-based candidate generation, deep learning-based semantic predictions, and an instance-level logistic regression model for classification of candidate pole instances. The method is evaluated against standalone geometric and deep learning baselines under a common source-to-target evaluation setting. The results show that the fusion approach achieves the highest F1-score in both source and target domains. In the target domain, it improves precision by more than 30 percentage points relative to the deep learning baseline while maintaining high recall. Overall, the findings demonstrate that instance-level semantic-geometric fusion improves the reliability of pole detection under domain shift and is a promising strategy for railway asset mapping when labeled target-domain data are scarce.