Keywords

Infrared imagery; Moving target detection; Automatic target recognition; Open-set recognition; Image registration; Inertial navigation systems

Abstract

Automated infrared (IR) imagery analysis is essential for persistent surveillance, defense, and security, yet it remains difficult when sensors move and when unknown objects appear. This dissertation addresses two fundamental problems: (1) detecting small moving targets amid platform-motion induced parallax and (2) distinguishing between known stationary targets and out-of-distribution (OOD) objects.   For detecting moving targets while the platform itself is in motion, auxiliary Global Positioning System and Inertial Navigation System data are combined with image analysis. Direction Cosine Matrices from calibrated inertial measurements enable sub-pixel frame alignment unattainable with purely image-based registration. The stabilized sequence is processed by a Reed–Xiaoli Anomaly (RXA) detector, with optical-flow suppression of residual parallax. In highly dynamic scenes, spatial Mexican-Hat filtering, motion-derived pseudo-stereo ranging, and a K-Nearest-Neighbors classifier further reduce false alarms. Tests on real mid-wave IR sequences with stationary and moving sensors consistently surpass the Infrared Patch-Image model and other baseline algorithms.   For stationary-object recognition, classifiers must reject unknown objects instead of confusing them with known classes. We address this problem using Simultaneous Classification of Objects with Unknown Rejection (SCOUR), a Bayesian scheme that improves the ability of any existing classifier to reject unknown objects. A primary classifier supplies class probabilities, while a secondary network regresses class-conditional representations of the same known-class data. During inference, the class-wise confidence from the classifier are weighted by per-class normalized similarity scores to form a combined confidence metric. We show that a simple threshold can applied to this metric to effectively separate unknown objects from known classes. On diverse known/unknown splits of the public DSIAC MWIR dataset, SCOUR achieves state-of-the-art or competitive open-set performance while preserving accuracy on known classes.   Together, these inertially aided registration and Bayesian fusion techniques markedly strengthen automated IR surveillance, delivering robust moving-target detection under ego-motion and reliable rejection of unknown objects.

Completion Date

2026

Semester

Spring

Committee Chair

Abdolvand, Reza

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Document Type

Dissertation/Thesis

Identifier

DP0053097

Download

Share

COinS
 

Accessibility Statement

This item was created or digitized prior to April 24, 2027, or is a reproduction of legacy media created before that date. It is preserved in its original, unmodified state specifically for research, reference, or historical recordkeeping. In accordance with the ADA Title II Final Rule, the University Libraries provides accessible versions of archival materials upon request. To request an accommodation for this item, please submit an accessibility request form.