Monocular Deep Learning Pipeline for 3D Bird Collision Risk Assessment on Power Transmission Lines
Year: 2026
Type: article
Abstract: Bird collisions with transmission lines represent a major environmental challenge for wildlife conservation. Estimation of bird mortality from collisions remains particularly limited in areas where the site accessibility for carcass surveys is restricted or impractical. This study evaluates whether a single low-cost video camera, combined with state-of-the-art computer-vision techniques, can deliver reliable collision-risk metrics and thus address this shortfall. We introduce the ARA–01 and ARA–02 datasets, the first specifically designed for detecting small, fast-moving birds in complex riverine environments. Three-dimensional flight paths were reconstructed by integrating a YOLOv11-L model (mean Average Precision at IoU = 0.5 [[email protected]] = 0.799; F1-score = 0.71), monocular depth maps from Depth Anything V2, and frame-wise Random Forest regression anchored to spherical markers. This approach achieved a Root Mean Square Error (RMSE) of 22.38 m and coefficient of determination (R2) = 0.929 across the 50–330 m range. Field trials on 10 and 14 October 2024 recorded 32 flights, of which only seven approached within 10 m of the wires. With 95 % confidence, none entered the <5 m high-risk corridor, though the small sample size and limited temporal window preclude conclusions regarding avoidance behavior or deterrent efficacy; further long-term monitoring is required before any causal interpretation. The system therefore provides a practical low-cost monocular setup capable of automatically generating risk indicators.
Authors: Pedro Lopes Ismerim Medina, Marco Antonio Ruiz Rueda, Benito Rafael Santana de la Torre, Ícaro Menezes Pinto, Pedro Ricardo Nery Nunes Silva, Gabriela Gama da Silva Santos, Viviane Spencer Andrade, Tacito Henrique da Silva Graça, Luiggi Cavalcanti Pessôa, Alessandra Schwertner Hoffmann, Michele Ferreira Lima, Ricardo Abranches Felix Cardoso Junior, Beatriz Cabral Dias de Carvalho
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7223703816