@ University of Southern California
In this work we consider the problem of tracking objects from a moving airborne platform through long occlusions and/or when their motion is unpredictable. The main idea is to take advantage of the known 3D scene structure to estimate a dynamic occlusion map and link fragmented tracks by solving a sequence alignment problem.
Here we developed an algorithm for tracking of vehicles in aerial surveillance imagery. The main idea is to formulate the data association problem as inference in a set of Bayesian networks, which offers a number of advantages. Results show low track fragmentation and high computational efficiency.
To classify vehicles from arbitrary viewpoint in video, we propose to estimate the moving vehicle pose using structure from motion and match a database of 3D models projected to the same pose.
@ University of Central Florida (graduate)
Head and Shoulder Detection
This project developed an algorithm for head detection. New low-level and mid-level features, optimized for curved object detection, are introduced to function in environments with much background clutter. Object detection is accomplished by transforming the mid-level features into weak classifiers and boosting.
This project developed a method for adding color to Picbreeder. Picbreeder is an application allowing users to generate art by interactively evolving neural networks. Originally, the algorithm used by Picbreeder was only capable of producing grayscale output. This project made it possible to add color to existing grayscale images as well as to evolve new color images from scratch.
@ University of Central Florida (undergraduate)
This project developed an algorithm to detect hands in a single image without relying on appearance modeling / skin detection. A new scale space edge detector finds scale invariant features at object boundaries. To detect a hand, a pattern of curves for finger tips and wedges, and lines, for finger sides, is defined. This pattern is identified in images using a graph based algorithm.
This project developed a numerical technique for quickly solving the one-dimensional diffusion equation. By spatially dividing the diffusion system into regions and using an appropriate time time step size in each region, calculation greatly speeds up with minimal loss in accuracy.