Date of Award
Doctor of Philosophy (PhD)
Computer and Information Science
The simplest Computer Vision algorithm can tell you what color it sees when you point it at an object, but asking that computer what it is looking at is a much harder problem. Camera and LiDAR (Light Detection And Ranging) sensors generally provide streams pixel of values and sophisticated algorithms must be engineered to recognize objects or the environment. There has been significant effort expended by the computer vision community on recognizing objects in color images; however, LiDAR sensors, which sense depth values for pixels instead of color, have been studied less. Recently we have seen a renewed interest in depth data with the democratization provided by consumer depth cameras. Detecting objects in depth data is more challenging in some ways because of the lack of texture and increased complexity of processing unordered point sets. We present three systems that contribute to solving the object recognition problem from the LiDAR perspective. They are: calibration, registration, and object recognition systems. We propose a novel calibration system that works with both line and raster based LiDAR sensors, and calibrates them with respect to image cameras. Our system can be extended to calibrate LiDAR sensors that do not give intensity information. We demonstrate a novel system that produces registrations between different LiDAR scans by transforming the input point cloud into a Constellation Extended Gaussian Image (CEGI) and then uses this CEGI to estimate the rotational alignment of the scans independently. Finally we present a method for object recognition which uses local (Spin Images) and global (CEGI) information to recognize cars in a large urban dataset. We present real world results from these three systems. Compelling experiments show that object recognition systems can gain much information using only 3D geometry. There are many object recognition and navigation algorithms that work on images; the work we propose in this thesis is more complimentary to those image based methods than competitive. This is an important step along the way to more intelligent robots.
Patterson IV, Alexander Evans, "Registration and Recognition in 3D" (2014). Publicly Accessible Penn Dissertations. 1401.