DLVS: Time Series Architecture for Image-Based Visual Servoing

A novel deep learning-based visual servoing architecture “DLVS” is proposed for control of an unmanned aerial vehicle (UAV) capable of quasi-stationary flight with a camera mounted under the vehicle to track a target consisting of a finite set of stationary points lying in a plane. Current Deep Learning and Reinforcement Learning (RL) based end-to-end servoing approaches rely on training convolutional neural networks using color images with known camera poses to learn the visual features in the environment suitable for servoing tasks. This approach limits the application of the network to available environments where the dataset was collected. Moreover, we cannot deploy such networks on the low-power computers present onboard the UAV. The proposed solution employs a time series architecture to learn temporal data from sequential values to output the control cues to the flight controller. The low computational complexity and flexibility of the DLVS architecture ensure real-time onboard tracking for virtually any target. The algorithm was thoroughly validated in real-life environments and outperformed the current state-of-the-art in terms of time efficiency and accuracy.