Research · BMVC 2026 · Accepted · Poster
Robust Utility Networks Registration with Powerline Segmentation on Noisy Point Clouds
Turning aerial LiDAR surveys into per-pole, per-conductor utility records, robust to badly classified point clouds.
- arXiv link coming soon
We wanted to turn free public USGS LiDAR into usable utility-asset data: locating poles, wires, and the network connecting them. But the data is sparse, vegetation-heavy, and messy, with no single end-to-end approach that reliably handles the problem.
Abstract
Aerial LiDAR has become the standard sensor for surveying overhead utility networks, yet converting a survey into a structured per-conductor digital asset record remains largely manual and time consuming. We present an end-to-end framework that converts aerial LiDAR surveys, raw or pre-classified, into structured, per-pole and per-conductor digital utility asset records. Network topology is reconstructed in a single forward pass by a dual-head ResUNet-FPN over four rasterised DSM channels, two class-dependent and two class-agnostic, regularised by ClassificationChannelDropout, a semantically structured masking scheme that forces the encoder to commit to the classification-invariant pathway and lifts pole detection by +7.1% (vs. the same model without the mask) on degraded inputs at no measurable cost on clean ones. For each reconstructed span, individual conductors are recovered by combining a PointNet-style learned segmenter trained entirely on synthetic spans with a training-free geometric pipeline (PCA alignment, catenary straightening, HDBSCAN), routed by a mixture-of-experts gate. Continuous catenary fits over the resulting per-point assignments make it feasible to extract endpoint heights, sag, clearance, length and inter-wire spacing as one-line queries. Across aerial LiDAR corpora spanning 0.4–193 pts/m² and clean-to-degraded classification, the system delivers PDet ≥ 0.91 (pole detection IoU) and WCLD ≥ 0.92 (wire detection score) on topology, 99.8% per-point precision on conductor assignment under consensus gating, and 12–18 cm MAE on endpoint and sag attributes.
Accepted as a poster at BMVC 2026, Lancaster, 23–26 November. Pictures from the conference to follow.
Anuja
First day in Washington, DC and BMVC decided to make it even better. Paper accepted to the BMVC 2026 Main Track. 🙂
View on X ↗
The decision, as it arrived.