Real-time GP-based wheelchair corridor following

Abstract

In this paper, we present a novel GP-based visual controller. The HOG features are used as a global representation of the observed image. The Gaussian Processes (GP) algorithm is trained to learn the mapping from the HOG feature vector onto the velocity variables. The GP training is achieved using corridor images collected from different places, these images are labeled using velocity values generated by a geometric-based control law and robust features. A hand-based verification of the features is done to ensure the accuracy of the ground truth labels. Experiments were conducted to explore the capabilities of the developed approach. Results have shown R Squared metric with more than ninety percent on the trained GP model in noisy conditions. © 2021 IEEE.