Rensselaer’s SEAL uses images for a variety of applications in space science and spacecraft navigation. The unique requirements of spaceflight often provide motivation for novel contributions in the area of image processing.
Subpixel Edge Localization
Edge localization (or edge detection) is a critical low-level image processing task used for a variety of image processing applications --- from finding the lit limb of a celestial body to processing medical images. Most classic edge detection algorithms (e.g. Sobel, Prewitt, Canny) are pixel-level and only report the integer pixel location of edges. Therefore, image pixelation becomes the limiting factor in edge localization accuracy for these methods. Many scientific, engineering, and medical applications can benefit from refining these pixel-level estimates to achieve subpixel-level accuracy.
Our subpixel edge localization algorithm uses Zernike moments to improve pixel-level guesses of an edge’s location. We improve on prior moment-based methods by replacing the underlying model of an edge with a ramp function (instead of a step function). The use of a ramp allows us to capture the effect of pixelation and image blur/defocus and eliminate biases present in past methods. The algorithm is simple to implement and is very fast.
Three-Dimensional (3D) Feature Descriptors
Technological advancements over the last 25 years have made 3D sensors inexpensive and widely available. Common examples of commercially available 3D sensing solutions include stereovision systems and Time-of-Flight (ToF) cameras.
These sensors produce a 3D point cloud of the observed scene. It is often important to align (i.e. register) an observed point cloud with the reference model. This has applications in both 3D object recognition and relative navigation.
The Rensselaer SEAL team is exploring methods for generating robust 3D feature descriptors. Our current work focuses on the development of a formal 3D scale space and on 3D feature histograms.