Visual Pursuit Control based on Gaussian Processes with Switched Motion Trajectories

TL;DR

This paper introduces a pursuit control method using switched Gaussian Process models for visual tracking of targets with changing behaviors, ensuring bounded errors and demonstrated effectiveness through simulation.

Contribution

It presents an improved Visual Motion Observer with switched GP models and an online switching behavior estimation method for dynamic target pursuit.

Findings

Bounded control and estimation errors with high probability

Effective switching behavior estimation demonstrated in simulation

Applicability to real-world scenarios confirmed through Digital Twin simulation

Abstract

This paper considers a scenario of pursuing a moving target that may switch behaviors due to external factors in a dynamic environment by motion estimation using visual sensors. First, we present an improved Visual Motion Observer with switched Gaussian Process models for an extended class of target motion profiles. We then propose a pursuit control law with an online method to estimate the switching behavior of the target by the GP model uncertainty. Next, we prove ultimate boundedness of the control and estimation errors for the switch in target behavior with high probability. Finally, a Digital Twin simulation demonstrates the effectiveness of the proposed switching estimation and control law to prove applicability to real world scenarios.