Date of Award

3-2025

Document Type

Thesis

Degree Name

Master of Science

Department

Department of Operational Sciences

First Advisor

Bruce A. Cox, PhD

Abstract

Classification “flickering,” where the classification of an object changes inconsistently between consecutive video frames, remains a persistent issue in modern object classification algorithms. This problem undermines the reliability of autonomous vision systems and poses significant risks in high-stakes applications such as autonomous vehicles. This thesis explores the use of response surface methodology, a statistical design of experiments technique, to optimize hyperparameters across three object classification pipelines. The first pipeline combines YOLOv8 with SORT to establish a benchmark. The second integrates a Bayesian back-end, while the third employs an exponential smoothing back-end. Hyperparameter tuning was conducted using a two-step process: an initial screening design to identify key factors, followed by central composite designs to refine hyperparameter settings and optimize multiobject tracking accuracy (MOTA). Results indicate that the exponential smoothing back-end achieves the highest MOTA at the cost of temporal stability, whereas the Bayesian back-end offers enhanced classification consistency across varied environments but at the cost of reduced MOTA. The YOLO-SORT baseline achieves balanced performance, with relatively high MOTA and moderate variance. These findings advance the development of autonomous vision systems by addressing critical trade-offs between stability and accuracy, with significant implications for improving safety in high-assurance applications.

AFIT Designator

AFIT-ENS-MS-25-M-164

Comments

An embargo was observed for posting this work.

Distribution Statement A: Distribution Unlimited. Approved for public release. PA case number: 88ABW-2025-0258

Download

Share

COinS