Uso de redes neurais profundas para detecção de vegetação em contato com linhas de distribuição de energia em ambientes urbanos

Abstract

Ensuring the reliability and safety of electric power distribution is essential for modern urban management. This research focuses on the development of a system to detect contact between vegetation and power distribution lines using deep neural network architectures, particularly recent versions from the You Only Look Once (YOLO) family and the Real-Time Detection Transformer (RT-DETR). The research methodology involved data collection, annotation, image pre-processing, model training, and performance evaluation. A low-cost mobile mapping system equipped with cameras mounted on ground vehicles was used to capture images, which were then labeled to identify vegetation in contact with distribution cables. Then, several image pre-processing techniques, including classical edge detection filters such as Canny, Sobel, and Prewitt, as well as other image filtering methods, were tested to assess their impact on detection quality. Among the architectures evaluated, YOLOv10m achieved the best performance, with an F1-score of 89.7% and a mAP of 92.7%. The proposed system aims to enhance the detection of hazardous conditions, reduce operational costs, and improve safety by providing a reliable tool for preventive maintenance in power distribution networks. Furthermore, the study publicly releases the labeled dataset to support and encourage future research.