Advancing Utility Pole and Sign Detection Through Deep Learning

Authors: Carl Dickinson, Gaetano Di Caterina

Submitted: August 4, 2026

Published in: Proceedings of the 36th British Machine Vision Conference (BMVC 2025), BMVA, Paper 976, 2025

DOI: 10.48550/arXiv.2608.04061

Abstract

Utility poles form a critical backbone of power distribution networks and other essential public services. Regular inspection of these structures is vital for maintaining the stability and safety of the electrical grid. This paper presents a deep learning framework for the automated detection, segmentation, and lean angle estimation of wooden utility poles, alongside the classification of attached electrical warning signs, using ground-level imagery.

The system was trained on a custom dataset of 4,570 annotated images sourced from Google Street View, featuring challenging real-world scenes with visually ambiguous wooden poles that lack distinctive features. The proposed model builds on the Detection Transformer (DETR) architecture, which was suitably modified and fine-tuned on this dataset.

Our model outperforms standard object detectors—including RetinaNet, Faster R-CNN, and YOLOv3-Tiny—achieving a mean average precision (mAP) of 90.43% for pole detection and 88.26% for sign detection. By extending the model with a segmentation head, we enable per-instance mask generation, which is then used to estimate the pole lean angle. The model accurately estimates lean for 1,367 out of 1,433 test-set poles, with a mean absolute error of just 1.01 degrees.

The custom dataset created for this work has been made publicly available, providing a benchmark for future research in utility infrastructure monitoring and related computer vision applications.

Introduction

The reliable operation of power distribution networks depends heavily on the structural integrity of utility poles. These wooden structures support overhead power lines, transformers, and other critical equipment, making their regular inspection a high-priority task for utility companies. Traditional manual inspections are labor-intensive, costly, and often impractical across vast service areas. As of 2026, the integration of automated computer vision systems into infrastructure maintenance workflows has become increasingly feasible, driven by advances in deep learning and the availability of large-scale street-level imagery.

This work addresses three key challenges in automated utility pole inspection: (1) detecting poles in cluttered, real-world environments, (2) segmenting them for detailed analysis, and (3) estimating their lean angle—a critical indicator of structural health. Additionally, we classify attached electrical warning signs, which provide essential safety information and regulatory compliance data.

Methodology

Dataset Construction

We compiled a custom dataset of 4,570 images extracted from Google Street View, capturing diverse geographic regions, weather conditions, and lighting scenarios. Each image was manually annotated with bounding boxes for poles and signs, per-instance segmentation masks, and lean angle ground truth. The dataset includes challenging cases where wooden poles blend into natural backgrounds or are partially occluded by vegetation, structures, or signage.

Model Architecture

Our approach is built on the Detection Transformer (DETR), a transformer-based object detection architecture that eliminates the need for hand-crafted components like anchor generation and non-maximum suppression. DETR formulates detection as a set prediction problem, making it well-suited for end-to-end training. We modified the architecture to incorporate a segmentation head, enabling simultaneous detection and per-instance mask generation. This dual capability allows lean angle estimation directly from the segmentation output.

The model was trained on the custom dataset with carefully tuned hyperparameters to balance detection accuracy and computational efficiency.

Results

Detection Performance

Table 1 summarizes the detection performance of our DETR-based model compared to traditional detectors.

| Model | Pole Detection mAP (%) | Sign Detection mAP (%) |

|--------------------------|------------------------|------------------------|

| RetinaNet | — | — |

| Faster R-CNN | — | — |

| YOLOv3-Tiny | — | — |

| DETR (Ours) | 90.43 | 88.26 |

Full comparative results are presented in the paper.

The DETR-based model achieves a mean average precision of 90.43% for pole detection and 88.26% for sign detection, surpassing all baseline detectors. Its transformer-based attention mechanism proves particularly effective at distinguishing visually ambiguous wooden poles from similar-looking vertical structures in the environment.

Lean Angle Estimation

Using the segmentation masks, we computed the lean angle for each detected pole relative to the vertical axis. The model successfully estimated lean angles for 1,367 of 1,433 test-set poles (a 95.4% success rate), with a mean absolute error of 1.01 degrees. This level of accuracy is sufficient for identifying poles that require corrective maintenance, as industry guidelines typically mandate action for lean angles exceeding 2–3 degrees.

Discussion and Future Work

Our results demonstrate that transformer-based detection architectures are highly effective for utility infrastructure monitoring tasks. The combination of detection, segmentation, and lean angle estimation in a unified framework simplifies the inspection pipeline and reduces the need for multiple specialized models.

The public release of our dataset and annotations provides a valuable benchmark for the research community, encouraging further innovation in this domain. Future work could explore:

• Extension to metal and composite utility poles
• Integration with aerial imagery from drones or satellites
• Real-time deployment on edge devices for field inspections
• Incorporation of temporal data to track pole deterioration over time

Conclusion

This paper presents a deep learning framework for automated utility pole and sign detection using a modified DETR architecture. The model achieves state-of-the-art performance on a challenging real-world dataset, with high detection accuracy and precise lean angle estimation. By making our dataset publicly available, we aim to accelerate research in infrastructure monitoring and support the modernization of electrical grid maintenance practices.

via ArXiv CV