Open Collections

Natural illumination invariant imaging

University of British Columbia

Shadows confound many machine vision algorithms and strong ambient illumination confounds active imaging, especially in outdoor automation applications such as surface mining. Natural Illumination Invariant Imaging takes images that appear to be illuminated by only an intentional flash, even with a flash only 1/37000th the power of the ambient light. This work combines hardware techniques, including brief pulses and filtering to reduce the apparent intensity of ambient light, with subtracting a reference image of the scene under only ambient illumination. Most existing techniques simply subtract an image of the scene taken at a previous time, or use several cameras to estimate the reference image. This work explores creating an accurate reference image using two cameras. The improved reference image for subtraction allows ambient shadows to be substantially removed from an image. Four experimental methods: reference estimation, segmentation, shadow removal and depth correlation; were used to evaluate the performance of ten techniques. Estimates of flash-free reference images created with the proposed techniques were compared to real images taken without a flash showed up to 63% reduction in error over existing techniques. The techniques were applied to three image processing methods: segmentation, shadow removal and depth correlation. They were applied to video for segmentation, and demonstrated segmenting an object 55% more accurately, though the measurement is scene dependent. Next, shadow ratio experiments examined the magnitude of the shadow remaining after subtraction, showing up to 70% shadow intensity removal. Depth correlation experiments showed over a 60% increase in the image area recognized using an efficient logical correlation algorithm. The theoretical performance of the techniques was examined. In summary, illumination invariant imaging was explored theoretically and experimentally, showing significant improvements for some important methods.

Text; Moving Image

2012-04-30

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/33741

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/