Remotely-Processed Vision-Based Control of Autonomous Lighter-Than-Air UAVs With Real-Time Constraints
Caeden Taylor, Matthew Widjaja, Or D. Dantsker
Abstract
This paper presents a remote computer vision processing system for autonomous lighter-than-air (LTA) vehicles that overcomes the computational limitations of onboard processing. The system utilizes a Raspberry Pi Zero 2W as the remote client and a high-performance computer as the remote processing server. The system achieves real-time object detection and control by transmitting images wirelessly to the server, which processes the images using a YOLOv8 model and returns the results to the client. The system is designed to balance detection accuracy, processing delay, and control responsiveness, and is evaluated in the context of the Defend The Republic competition. The results show that the system can detect objects of interest at over 10 frames per second, primarily limited by the communication network's bandwidth. The system's performance is analyzed in terms of image resolution, model size, network bandwidth, and image compression, and the trade-offs between these factors are discussed. The paper concludes that remote computer vision processing is a viable solution for autonomous LTAs, enabling the use of more advanced computer vision models and improving system performance.