A Concept for Increasing Trustworthiness in Deep Learning Perception for UAS Using Map Data

Kurzfassung

Deep Learning (DL)-based perception models provide state-of-the-art results in semantic segmentation and object detection, allowing an unmanned aircraft system (UAS) to perceive and understand its environment. This capability is particularly crucial for applications like onboard risk-based planning, where a UAS dynamically adjusts its trajectory given critical changes that are detected during flight. However, fully trusting the output of a DL model without safeguards is inadvisable, as DL models are regarded as black boxes, lacking explainability and interpretability of their outputs. To enhance trust in the DL output, regulatory bodies recommend monitoring the model output during flight by an independent system. In this work, we propose a runtime monitoring concept for a DL-based UAS environment perception system that detects static and dynamic objects. We increase the trustworthiness of the system by computing the plausibility of the model output using trustworthy map data and known contextual relationships. For static objects such as buildings and streets, the plausibility is calculated against the map data. The plausibility of dynamic objects such as pedestrians or vehicles, which are not present in the map data, is calculated using an ontology derived from map data. After presenting the concept as well as its strengths and weaknesses, future avenues for deploying the concept are highlighted.