Airport runway detection using convolutional neural networks

Abstract

The recent advancement in computational power allowed the world to give more use of deep learning algorithms. The image processing task is on itself data-hungry, but showed to be of exetreme importance in many fields of knowledge. In this work, a runway detection method, for aircraft during landing phase, based on convolutional neural network is made with the motivation to reduce pilot workload, increase flight safety during landing and take new steps towards single-piloted or even unmanned commercial flights. With the results, the network could successsfully converge its loss function on the training dataset and alidation dataset, and inference on severeal images from test dataset and real runway images gave good results for keypoint and bounding boxes predictions. Our numerical results indicate that virtual runway images can be used in order to pretrain a netwaork to detect real runway images, so we contribute to a chearper and faster approach on the development of suck technology. Further analysis on estimatives for orientation and position of the aircraft camera from predicted keypoint indicate the work to be promising. Even being an initial work, the results gave a robust error histogram which could be easyly controlled by filtering outliers and by data fusion with other sensors in real application. Overall, we contribute with results indicating that virtual training, using data augmentation methods for dataset enrichment, benefits real detection and that the keypoint predictions may be used together with pose estimation algorithm to give estimates of aircraft pose related to runway.