Towards Visual Navigation Input Monitoring in Autonomous Aerial Vehicles

Autonomous aircrafts have by design to master flights without pilot interventions. Intelligent perception/decisionmaking capabilities accordingly empower the unmanned system to cope with unforeseen situations. In particular, the safe deployment of pilotless vehicles providing services in the vicinity of critical infrastructure represents a considerable challenge. In order to comply with regulatory requirements, the augmentation of aerial vehicles with dedicated components might appropriately enforce robust autonomy. Measurement technology may mitigate some safety and compliance risks: for enhanced reliability an additional component should namely evaluate sensor data before it is ingested by an aircraft’s intelligent navigation. Notably here, but also more broadly, metrology is increasingly faced with big, high-dimensional and dynamic data. In this paper, the monitoring of visual navigation input is therefore explored. First, obstacle avoidance is dissected, as a first exemplary computer-vision task to be safeguarded. Second, nadir-pointing visual odometry is analogously examined as a common localization approach. Considering both those relevant cases, recent advancements in deep image understanding as well as dimensionality reduction are studied with respect to computerized monitoring of incoming optical data. An aim of the analysis consists in the detection/rejection of adversarial or out-of-scope input. Precisely, the contributions of this present paper include: (i) adding undisclosed details about the public ISeVis dataset, (ii) combining powerful general-purpose vision encoders with flexible low-dimensional representation for comprehensible information extraction from input imagery, (iii) demonstrating capabilities and limitations of the approach drawing on two open datasets. Lastly, future development directions are described.