Short URL

Parallelization of Computer Vision Techniques for Large-Scale Services

 

Other areas:

Loading, please wait...

Project Overview

This project funds two CSCE computer science students to investigate how high-performance computing resources can be used to improve selected geospatial methods. As such it draws on multiple and disparate research areas including large-scale systems, distributed and parallel programming, photogrammetry, and computer vision. The goal of this research is to leverage large-scale systems, especially distributed storage and computation, to provide advanced geospatial image services to end-users. The research will leverage large-scale storage to hold multiple terabytes of aerial and satellite imagery and will advance understanding across multiple fields of research by relating the problems in one discipline to other disciplines. This activity will foster collaboration between researchers in geomatics, computer vision, and parallel and distributed computing.

The problem under investigation is the autonomous orientation of large amounts of aerial/satellite imagery.  The information generated is used in studies, for example examining the changing path of rivers or floodplains, city growth, and deforestation. In each of these cases, large geographic areas are examined. This project uses parallel and distributed processing in conjunction with computer vision techniques to find and extract “interesting” and unique features in images. Once processed, photogrammetric methods orient (via a robust bundle adjustment) and orthorectify the images for integration into geographic information systems. In previous work, we have focused on a class of methods known as affine invariant feature detectors and descriptors. Two students parallelized and accelerated the sift++ application which is an implementation of the Scale Invariant Feature Transform (SIFT) algorithm. The students' work examined the speedup obtainable with single node parallelization and GPGPU acceleration with NVIDIA's CUDA framework.

 

Results

The results of this research will be accessible to users across the state and be used to teach graduate, undergraduate, and high-school students about photogrammetry, computer vision, and parallel and distributed computing. The multiple levels of depth of understanding allow this kind of research to be tailored to semester long graduate courses, or to an hour-long demonstration in a high-school. The techniques used in this work meld several areas of research that can be mirrored in a cross-disciplinary course that could, for example, pair computer science, mathematical science, and geography students to complete course projects that require data processing, algorithm development, and GIS processing. Additionally, each of these facets can be separated into individual courses. The front-end can be used to teach information systems students how to develop scalable services that rely on asynchronous processing of data, and use large database systems to service needs.

 

Sponsor

National Science Foundation (NSF)

 

Collaborator

Amy Apon

 

P.I.

Jackson Cothren