Uncertainty Quantification-Based Switching Control Method for Vision-Based Object Tracking in Unmanned Aerial Vehicles

Loading...
Thumbnail Image

Authors

Fernández Castaño, Antonio

Issue Date

2024

Type

Thesis

Language

Keywords

Machine Learning , Object Detection , Switching Controller , Unmanned Aerial Vehicles

Research Projects

Organizational Units

Journal Issue

Alternative Title

Abstract

Ensuring safe aerial robotics has become paramount with the increasing utilization of UAVs (Unmanned Aerial Vehicles) in dense urban environments. Machine Learning-based object detection methods that process camera sensor images have become essential in planning and navigating UAVs through these environments. However, like any machine learning tool, the object detection model is prone to aleatoric and epistemic uncertainties. This thesis introduces a perception-control framework incorporating a switching controller to improve robustness against uncertain measurements processed by an object detector. The proposed mechanism enables switching between different controllers based on the level of uncertainty quantified. This approach aims to enhance the reliability and safety of UAV operations in challenging environments. We tested this framework using a simulated environment where a UAV tracks an object using the YOLOv5 object detector. Our results demonstrate that the proposed control framework can mitigate the uncertainties of the machine learning perception by decreasing the system gain from the sensor noise to the output of the vehicle dynamics.

Description

Citation

Publisher

License

Creative Commons Attribution-NonCommercial 4.0 International

Journal

Volume

Issue

PubMed ID

DOI

ISSN

EISSN