3D keypoint detection for robust range image registration and scene understanding. We are working on novel algorithms to detect interesting features in range data, with the goal of constructing detailed 3D models of urban scenes by combining range and color image data.
Hole Feature Extraction
Planar Segmentation and Hole Extraction
My current work deals with extracting holes from dense range scans. By describing holes via spin images, we are able to a.) get a very good initial transformation for registration and b.) perform object recognition by indexing.
Spin Images applied to windows of the VCC.
Spin Images applied to windows of the biotech building. (SCAN 1)
Spin Images applied to windows of the biotech building. (SCAN 2)
Refinement via ICP
View #1 of the VCC.
View #2 of the VCC.
We use the Iterative Closes Point Algorithm to refine the initial transformation derived from a single feature match.
Registered Views.
Transformation Selection
After several transformation have been refined, we use free space constraints and algorithms that we call re-spinning and re-keying to find the single best transformation.
Publications
B. King, T. Malisiewicz, C. Stewart, R. Radke, Registration of Multiple Scans as a Location Recognition Problem: Hypothesis Generation, Refinement and Verification . 3DIM 2005: 180-187.
Researchers
The principal investigators in this research are: Tomasz Malisiewicz (left), Brad King (right), Dr. Chuck Stewart (center), Dr. Rich Radke (not pictured), and a Leica range scanner.