Time Series Approach for Visual Servoing Using Transformers

We introduce Visual servoing using a novel deep-learning time-series architecture to control an unmanned aerial vehicle (UAV) with a mounted camera to track a target consisting of a finite set of stationary points lying in a plane. Many visual servoing approaches use computer vision along with estimation algorithms, sensors, and actuators' feedback to solve tasks like, tracking, obstacle avoidance, and localization. Nowadays, deep neural networks are gaining popularity in such tasks owing to their accuracy, adaptability, and flexibility. We propose a solution that employs a time-series architecture to learn temporal data from sequential values to output the control cues to the flight controller. Because of its low computational expense, the solution is deployable on less powerful onboard computers present on the UAV, ensuring real-time tracking of the target. The solution is tested both in a simulation environment and in real life, outperforming the current state-of-the-art in terms of time efficiency and accuracy.