Deep Learning Approach for UAV Visual Electrical Assets Inspection

• Autor(a) principal: Barbosa, Joel Adão Pacheco
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.22/16898
• topic: Electrical Assets Inspection; Deep Learning; Object Detection; Multi-Object Tracking; UAV; Inspeção de Ativos Elétricos; Deteção de Objetos; Seguimento de Múltiplos Objetos; VANT

The growth in the electrical demand by most countries around the world requires bigger and more complex energy systems, which leads to the requirement of having even more monitoring, inspection and maintenance of those systems. To respond to this need, inspection methods based on Unmanned Aerial Vehicles (UAV) have emerged which, when equipped with the appropriate sensors, allow a greater reduction of costs and risks and an increase in efficiency and effectiveness compared to traditional methods, such as inspection with foot patrols or helicopter-assisted. To make the inspection process more autonomous and reliable, most of the methods apply visual detection methods that use highly complex Deep Learning based algorithms and that require a very large computational power. This dissertation intends to present a system for inspection of electrical assets, able to be integrated onboard the UAV, based on Deep Learning, which allows to collect visual samples grouped and aggregated for each electrical asset detected. To this end, a perception system capable of detecting electrical insulators or structures, such as poles or transmission towers, was developed, using the Movidius Neural Compute Stick portable platform that is capable of processing lightweight object detection Convolutional Neural Networks, allowing a modular, low-cost system that meets real-time processing requirements. In addition to this perception system, an electrical asset monitoring system has been implemented that allows tracking and mapping each asset throughout the inspection process, based on the previous system’s detections and a UAV navigation system. Finally, an autonomous inspection system is proposed, which consists of a set of trajectories that allow an efficient application of the monitoring system along a power line, through the mapping of structures and the gathering of insulator samples around that structure.