Degree Type

Creative Component

Semester of Graduation

Summer 2019

Department

Computer Science

First Major Professor

Dr. Rafael Radkowski

Second Major Professor

Dr. Jin Tian

Degree(s)

Master of Science (MS)

Major(s)

Computer Science

Abstract

The research addresses the visual calibration of head-mounted displays such as the HoloLens. The HoloLens is an optical see-through viewing device that allows a user to experience the real world populated with virtual objects. These virtual objects need to be correctly aligned with physical objects in the environment to experience a visually appropriate scene. However, several factors, such as an outside-in tracking system, tracking errors, the user's eye position, and others degrade the alignment between the virtual and physical object. A popular calibration method to correct this misalignment is the so-called Single Point Active Alignment Method (SPAAM) [1]. It allows one to improve the alignment by measuring and correcting the alignment error. Nonetheless, one encounters alignment errors since, SPAAM assumes a constant error between the physical object, the display, and the user's eye. Modern low-cost tracking systems such as based on RGB-D cameras (e.g., Kinect) come with dynamic errors. Consequently, SPAAM cannot yield the required accuracy; theoretically, dynamic errors require a dynamic calibration. The objective of this research is to study the improvement a dynamic error calibration can yield regarding alignment and registration accuracy. To improve the visual experience for a user, a random forest method will be adopted for this purpose. The hypothesis is that the random forest can dynamically select the best SPAAM calibration matrix with respect to the relative position of the user and a physical object. Experimental results demonstrate improvement by a factor of four; thus, indicate that random forest is an appropriate method to mitigate object misalignment due to dynamic tracking errors.

Copyright Owner

Sravya Kanuganti

File Format

PDF

Share

COinS