Project

DESCRIPTION:

Deep neural networks (DNNs) show promise in applications like agriculture, automotive, and construction machinery autonomy. DNNs can adapt to variations in lighting, background clutter, etc., making them robust in real-world scenarios where captured images are of a diverse quality due to different operating and environmental conditions. However, resource limitations such as computational power, memory, and energy consumption constrain their use on embedded devices. To address this, federated learning (FL) is proposed, where local models are trained on embedded devices, to be called below as clients, and aggregated on a central server to create a global model. This approach aims to overcome resource challenges in deploying DNNs on edge devices. This project aims to deliver a virtual testbed where a digital twin of an autonomous rover will be parametrized to account for possible variations in captured images from the rover’s cameras, and a FL framework will be integrated to enable tradeoff studies and help define the computational, memory, and energy requirements for the edge devices.