A FRAMEWORK FOR TESTING AND VERIFICATION OF VISION-BASED AUTOMATED LANDING

Jia, Yixuan

Permalink

https://hdl.handle.net/2142/124843 Copy

Description

Title

A FRAMEWORK FOR TESTING AND VERIFICATION OF VISION-BASED AUTOMATED LANDING

Author(s)

Jia, Yixuan

Issue Date

2023-05-01

Keyword(s)

safe autonomy; vision-based control

Abstract

Thanks to the recent progress of deep learning-based computer vision algorithms, we are now able to get accurate state and pose estimates directly from images. The complete perception pipeline usually consists of deep neural networks (DNNs) alongside other modules for applying traditional filters and transforms. The complexity of DNNs makes the verification of closed-loop systems consisting of such perception modules difficult. If we are to deploy such perception modules in real-world applications, especially in safe-critical tasks, some combination of testing, formal analysis, and understanding of the safe operating conditions have to be developed. In this thesis, we develop a testing framework for vision-based automated aircraft landing. The aircraft uses a convolutional neural network (CNN) to perform pose estimations, which is in turn used for navigation and control. We present the overall framework and describe the implementation of each component in detail. We then present end-to-end simulation results which can be used for testing and characterization of the safe operating conditions.

Type of Resource

text

Language

eng

Owning Collections