2003-2004

In my first real research position, I worked on a project in a biometrics lab at R&D on NSA's Fort Meade, MD campus. This project was interested in real-time facial recognition.

Our first step was to create a sample database of known subjects. A combination of 3-d scanning (using laser rangefinding) and a 360° array of cameras provided a set of 3-d meshes ready for import into a modeling/viewing suite. Filters (MatLab) were used to deal with potential inconsistencies with the data, and Autodesk's 3d Studio Max was used for texturing/manipulating/viewing the model.

Next, using 2-d video feeds (sequences of images) from "surveillance" cameras, a match was attempted between subjects in the feed and those in our "known subjects" database. Lighting conditions, facial augmentations (head hair, facial hair, jewelry, new scars, weight change, sunglasses, hats), angle of face, physical obstructions of the face, inconsistencies in 3-d mesh models, and poor resolution of the image sequence all presented problems with matching.

As an example method to deal with the problems above, we would try to match the lighting conditions in our 3-d world to those found in the 2-d imagery, create obstructions of our own to match those in the video feed, and rotate/scale/translate the facial mesh to match those found in the feed.

Images, video, and code cannot be linked due to the classified nature of this work. I've always wanted to say that.