Resumo:
High power power transmission lines are of great importance for the functioning
of all sectors of society. To guarantee the reliability and availability of the energy
supply, regular and occasional inspections are carried out, always manually. These
inspections seek to identify anomalies in the transmission network and signs of
danger to structures. Unmanned aerial vehicles, popularly known as drones, have
been increasingly employed in repetitive and tedious activities, due to their low cost
and increased safety, and can be applied to tracking these structures. However, the
variety of environments and towers, error in location measurements, and high cost
are difficult obstacles to overcome for accurate and quality inspection. The main
objective of this work is to study, develop and compare real-time image processing
techniques to determine movements to be performed by the drone in trajectory
correction, and thus perform a tracking of transmission lines with high precision
and low cost. Several solutions are found in the literature, and are located and
systematically reviewed in this study. Regarding the analysis of the hardware line,
controller boards and frameworks were compared, as well as ground station and
simulator software, with the main focus on image processing using edge detector,
deep learning and reinforcement learning. Furthermore, in relation to the training
of networks, a dataset was made for identifying lines, and another for actions to
be performed. Also, in order to evaluate the solution’s performance, a realistic
virtual scenario based on a real route was produced. In short, a system based on
a prototype equipped with two quadricopter drones equipped with only a camera
and a computer, working in conjunction with a cooperative algorithm and deep
learning image processing, resulted in a brief tracking in a power transmission line,
simulating an inspection in a real environment