Visual Servoing using FPGA-based Hardware accelerated Deep Learning Solution

A novel deep-learning time-series architecture is proposed for visual servoing to control an unmanned aerial vehicle (UAV). This UAV is capable of quasi-stationary flight and is equipped with a mounted camera to track a target composed of a finite set of stationary points located in a plane. The proposed solution utilizes a time-series deep learning architecture to learn temporal data from sequential values. It then generates control cues to be sent to the flight controller, enabling the UAV to perform its tracking task. We also present an optimized implementation of the novel proposed architecture on FPGA. When it comes to implementing Neural Networks, FPGAs have demonstrated their efficiency. Specifically, our implementation is quicker than top-tier central processing units (CPUs) when it comes to carrying out the inference of the suggested model. Moreover, the proposed system also achieved improvements in terms of energy efficiency, which is important for UAV’s longer endurance.

I. Introduction

Visual servoing refers to a control method for robotic maneuvers that utilizes optical data from present sensors. Its objective is to guide a robot from one point to another, through space whether specified or unspecified, by minimizing the error iteratively between its present position and the target position. There are several uses for visual servoing, such as robotic surgery and autonomous navigation [1].