Automated gross and sub-pixel registration accuracy of visible and thermal imagery

  • Authors:

  • Stephen Won;S. Susan Young;Gunasekaran Seetharaman

  • Affiliations:

  • Army Research Laboratory, Adelphi, MD;Army Research Laboratory, Adelphi, MD;Air Force Research Laboratory, Rome, NY

  • Venue:
  • Proceedings of the 10th Performance Metrics for Intelligent Systems Workshop
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

Tracking or improving resolution of images requires that objects of interests are aligned at every frame through registration. Three motion segmentation algorithms utilizing the image gradient and five motion estimation algorithms were tested to see which would produce the best registration process at the gross and subpixel level. To test the motion segmentation, a qualitative study was performed using visible and thermal imagery to see how accurately the area of motion could be identified without producing noise. Gross shift estimation accuracy was accessed by comparing estimated shifts against a ground truth. Sub-pixel shift estimation accuracy was assessed by a series of synthetically downsampled images. The initial evaluation of segmentation revealed that the flux tensor method performed the best with a fixed threshold. For gross-shift estimation accuracy on images with well-defined objects, feature correspondence yielded the most accurate results. For images with objects that are not well-defined, optical flow methods produce more accurate results. The sub-pixel shift estimation test revealed that correlation was the only method out of the four tested that could accurately estimate motion at the sub-pixel level.