Vision-Based Power Line Cables and Pylons Detection for Low Flying Aircraft

TL;DR

This work addresses the safety-critical problem of detecting distant ground obstacles—power line cables and pylons—for low-flying aircraft. It presents a single CNN with a ConvNeXt-Tiny backbone that jointly regresses distance maps for cables and pylons, trained with a composite loss that combines a distance-focused data term and a MALIS connectivity term, producing outputs in the range $[0,1]$ with $d_{max}=128$. Validated on the NE-VBW and the authors' DDLN datasets, the method outperforms prior distant-cable baselines and remains competitive for pylon detection, all while being deployed on an onboard flight system. The approach offers a practical, real-time solution that enhances pilot situational awareness by providing a reliable, second pair of eyes in challenging conditions.

Abstract

Power lines are dangerous for low-flying aircraft, especially in low-visibility conditions. Thus, a vision-based system able to analyze the aircraft's surroundings and to provide the pilots with a "second pair of eyes" can contribute to enhancing their safety. To this end, we have developed a deep learning approach to jointly detect power line cables and pylons from images captured at distances of several hundred meters by aircraft-mounted cameras. In doing so, we have combined a modern convolutional architecture with transfer learning and a loss function adapted to curvilinear structure delineation. We use a single network for both detection tasks and demonstrated its performance on two benchmarking datasets. We have integrated it within an onboard system and run it in flight, and have demonstrated with our experiments that it outperforms the prior distant cable detection method on both datasets, while also successfully detecting pylons, given their annotations are available for the data.

Figures (6)

• Figure 1: Architecture of the joint power line cables and pylons detection model. Output heads regress cables and pylons distance masks directly from the common feature map $F{}$ generated by the fine-tuned truncated ConvNeXt-Tiny feature extractor.
• Figure 2: A selection of two outputs for two loss function variants illustrating the application of $\mathcal{L}_{\texttt{MALIS}}$ connectivity loss term. The rows contain from left to right: an input image patch, a ground-truth cables segmentation mask, prediction of a model trained with MSD loss alone and prediction of a model trained with MSD loss combined with $\mathcal{L}_{\texttt{MALIS}}$ connectivity term. The two last images in each row contain distance masks in which $[0, 1]$ values range was mapped to a grayscale range from white to black. $\mathcal{L}_{\texttt{MALIS}}$ is applied to distance mask predictions within the marked square windows. Points A and B belong to two disjoint background regions, which should be separated by the cable detection in between. In the exemplary distance masks predicted for the cables with independent MSD loss supervision these points can be joined with a path traversing through not sufficiently small distance values. This can result in a broken cable detection in the final segmentation, i.e. in a thresholded distance mask. Such segmentation means having an unwanted merger of the background regions which $\mathcal{L}_{\texttt{MALIS}}$ aims to prevent.
• Figure 3: Examples of the images from the NE-VBW dataset Stambler19 (left) and from our DDLN dataset (right).
• Figure 4: Exemplary predictions of the three models trained on the NE-VBW dataset Stambler19: DeepWireCNN Stambler19, DDLN-S and DDLN-L. The inputs are $1536\times1536$ px crops from full-resolution images from the validation set. Predictions were generated by models initialized from checkpoints at epochs with highest validation cable segmentation quality metric.
• Figure 5: Exemplary predictions of the two size variants of our model trained jointly for cable and pylon detection on the DDLN dataset. The inputs are $1024\times1024$ px crops from full-resolution images from the validation set. Predictions were generated by models initialized from checkpoints at epochs with highest validation cable segmentation quality metric.