Date of Award


Level of Access Assigned by Author

Open-Access Thesis

Degree Name

Master of Science (MS)


Spatial Information Science and Engineering


Reinhard Moratz

Second Committee Member

Kate Beard

Third Committee Member

Nicholas Giudice


Handheld devices with depth sensors have the potential to aid low-vision users in performing tasks that are difficult with traditional modes of assistance. Heuristic studies have revealed that tables have a key functional role in indoor scene descriptions. The research question addressed in this thesis is: how can we robustly and efficiently detect tables in indoor office environments? This thesis presents a solution that utilizes a functional approach to robustly detect rectangular tables in depth images generated from a Kinect sensor.

Perhaps the most significant function of a table is to provide its users with a supporting plane. This demands that the table’s surface is orthogonal to the scene’s gravity vector. In order to fully take advantage of this functional property in the detection process, the scene must be properly oriented. A planar model fitting procedure is used to detect the scene’s floor, which is utilized to properly orient the scene.

The scene is then sliced at average table height, using a small buffer. The height component is removed from the 3-dimensional slice by projecting it into a two-dimensional plane. Next, an iterative labeling procedure is used to separate the image into independent blobs, allowing for 2-dimensional shape detection.

Sufficiently large blobs are then subjected to a cleaning process in order to remove any extraneous features. Several features of the cleaned blobs are calculated and used in a supervised classification process. The coordinates of blobs that are classified as tables are translated back to 3-dimensions, allowing for the segmentation of all detected tables in the scene.

Included in

Engineering Commons