Date of Award

3-14-2014

Document Type

Thesis

Degree Name

Master of Science

Department

Department of Electrical and Computer Engineering

First Advisor

Brian G. Woolley, PhD.

Abstract

In the eld of three-dimensional modeling, we continually struggle to quantify how closely the resulting model matches the physical object being represented. When precision measurements are required, they are often left to high-end, industrial systems. The aim of this thesis is to quantify the level of precision that can be obtained from commodity systems such as the Microsoft Kinect paired with the KinectFusion algorithm. Although the Kinect alone is considered a noisy sensor, the KinectFusion algorithm has shown the ability to build detailed surface models through the aggregation of depth information taken from multiple perspectives. This work represents the first rigorous validation of the three- dimensional modeling capabilities of the KinectFusion algorithm. One experiment is performed to measure the effects of key algorithm parameters such as resolution and range, while another is performed to measure the lower bounds at which objects can be detected and accurately modeled. The first experiment found that the KinectFusion algorithm reduced the uncertainty of the Kinect sensor alone from 10 mm to just 1.8 mm. Furthermore, the results of the second experiment demonstrate that the KinectFusion algorithm can detect surface deviations as little as 1.3 mm, but cannot accurately measure the deviation. Such results form an initial quantification of the KinectFusion algorithm, thus providing confidence about when and when not to utilize the KinectFusion algorithm for precision modeling. The hope is that this work will open the door for the algorithm to be used in real-world applications, such as alleviating the tedious visual surface inspections required for USAF aircraft.

AFIT Designator

AFIT-ENG-14-M-40

DTIC Accession Number

ADA599451

Share

COinS