Date of Award

3-11-2011

Document Type

Thesis

Degree Name

Master of Science

Department

Department of Electrical and Computer Engineering

First Advisor

Michael J. Mendenhall, PhD.

Abstract

This thesis provides a novel visualization method to analyze the impact that articulations in dismount pose and camera aspect angle have on histograms of oriented gradients (HOG) features and eventual detections. Insights from these relationships are used to identify limitations in a state of the art skin cued HOG dismount detector's ability to detect poses not in a standard upright stances. Improvements to detector performance are made by further leveraging available skin information, reducing false detections by an additional order of magnitude. In addition, a method is outlined for training supplemental support vector machines (SVMs) from computer generated data, for detecting a wider range of poses and camera configurations. The multi-SVM structure yields a 7-fold increase detection probability when applied to challenging crouching poses. These dramatic improvements clearly demonstrate the viability of such an approach, which can be extended to include other pose configurations.

AFIT Designator

AFIT-GE-ENG-11-05

DTIC Accession Number

ADA540101

Share

COinS