Deep Learning-based Utility Pole Safety Assessment from Visual Data

Utility poles are crucial infrastructure components, and efficiently assessing the safety of these structures and ensuring they adhere to the clearance guidelines, which specify the minimum distance between the pole and any surrounding objects, remains a challenge; the current manual inspection process is time-consuming, costly, and often subjective. This work proposes an automated deep learning-inspired solution to improve utility pole detection and measure the clearance distance. The biggest challenge was the lack of a comprehensive pole dataset; therefore, we collected a dataset containing utility poles in varied backgrounds, environments, and conditions. We compared data augmentation techniques and employed them to address the limited dataset size. The proposed approach consists of two main stages: pole detection and differentiation and pole distance measurement. The first stage is a comparison of multiple object detection models on our utility pole dataset; we used the results from the best-performing model to estimate the distance between the two pole objects. The results show that our pipeline with the YOLOv8 model outperforms SSD and achieves 83% accuracy in classifying pole compliance. The system can accurately detect and estimate clearance violations even with limited data. The success of the pipeline opens opportunities for future research; obtaining depth by using additional sensors or deep learning models could enhance the detection module. Scaling the approach to large utility pole networks while retaining real-time performance could lead to improved autonomous infrastructure maintenance.