Detecting cables and power lines in Small-UAS (Unmanned Aircraft Systems) images through deep learning

Small Unmanned Aircraft Systems (UAS’s) will be a new modal for cargo transport within urban and suburban areas and for security services, as they are small, agile and low-cost vehicles at lower altitudes. However, its use in urban areas has several challenges. A crucial one is the detection of cables and power lines on power poles that are exposed on streets, as this detection bring many other challenges to be faced, e.g., they are difficult to be detected because their structure is long and thin when compared to a noisy image background. Therefore, detection systems can avoid accidents that cause material damage and injury for people. One way to avoid accidents caused by this type of vehicles is through the detection of cables and power lines with an embedded camera installed in each small UAS. The proposed method on this work is to realize the semantic segmentation of cables and power lines in small-UAS images through computer vision applications and deep learning techniques. Convolutional neural network (CNN) UNet was customized and implemented, and UAS images dataset of power lines in urban and suburban areas was used to train, test and detection validation. The Adam optimizer algorithm was used in replacement for stochastic gradient descent and an accuracy of 0.96 was obtained, which can be considered a promising result to proceed with streaming videos. The main contribution of this research is to provide a detection of cables and power lines for a detect and avoid systems (DAA) of small UAS, which will turn safe the use of aircraft at lower altitudes in urban areas. In addition, a secondary contributing is to create a test standard in this area.