Performance Comparison of Deep Learning Networks for Runway Recognition in Small Edge Computing Environment

The purpose of this work is to investigate and compare three popular segmentation networks for runway recognition in a small edge computing environment that has the potential to become a part of avionics in the near future. For the networks, we examine YolactEdge, STDC-seg, and DDRNet, and for the small edge computing environment, we choose NVidia Jetson Orin AGX and Xavier NX. In order to verify their real-time performance as an on-board system, we first tested them on three different edge computers. We also conducted multiple flight tests with an actual aircraft and constructed a runway image dataset that was used for the training and evaluation of the networks. According to our result, YolactEdge, which is an instance segmentation model, showed a good performance, but it was difficult to run in real-time on our small computing environment. STDC-seg and DDRNet, which are semantic segmentation models, presented over 0.95 mIoU with approximately 0.1 seconds of inference time. In particular, DDRNet-23-slim, which is the smallest model of DDRNet, showed the highest mIoU and the shortest inference time among the networks we tested in this work.