Machine vision for UAS ground operations: using semantic segmentation with a bayesian network classifier

This paper discusses the machine vision element of a system designed to allow Unmanned Aerial System (UAS) to perform automated taxiing around civil aerodromes, with only a monocular camera. The purpose of the computer vision system is to provide direct sensor data which can be used to validate vehicle position, in addition to detecting potential collision risks. In practice, untrained clustering is used to segment the visual feed before descriptors of each cluster (primarily colour and texture) are used to estimate the class. As the competency of each individual estimate can vary dependent on multiple factors (number of pixels, lighting conditions and even surface type). A Bayesian network is used to perform probabilistic data fusion, in order to improve the classification results. This result is shown to perform accurate image segmentation in real-world conditions, providing information viable for localisation and obstacle detection.