Enabling Autonomous Unmanned Aerial Systems via Edge Computing

Unmanned Aerial Systems (UASs) have continuously demonstrated incredible value assisting with disasters such as wildfires and hurricanes. For example, UASs can help reduce risk in firefighting and increase useful data that can aid in developing a more informed strategy. Yet, performing tasks safely through tight spaces and accurately detecting nearby objects remains a major challenge facing fully autonomous flying. Due to the safety concern, CAL Fire has resisted the use of fire service UASs due to the unreliability of collision avoidance. Realizing the full potential of UASs for assisting with disasters will call for autonomous UASs that must be autonomous, taskable, and adaptive to incident situations, and respect safety, privacy, and regulatory concerns. In this paper, we propose the development of autonomous UASs capable of autonomous navigation, localization, 3-D mapping, and achieve on-board data processing and decision making. The UAS will fly and make decision using only on-board sensors and processors. Our contribution covers hardware design and embedded programming to multi-modal sensing, vision-based navigation, and hybrid mapping. We developed a new edge computing and sensing system for UASs which is compatible with existing open source autopilot software and deep-learning frameworks. We proposed a multi-modal sensing based hybrid localization and obstacle detection approach that runs in real time on board. The output of the localization and obstacle detection results is fused with high-level understanding and is used to control the UASs locally without rely on the link to a ground station. Our evaluation results demonstrate an autonomous UAS flying based on pre-defined destinations with on-board deep learning for perception and obstacle avoidance.