Using Particle Filters to Detect Polygon Structures of Fixed Shape in Still Images

D.W. Paglieroni and S. Manay (USA)


Pattern Recognition, Computer Vision, Overhead Imagery, Sequential Monte-Carlo


We introduce a novel method that uses particle filters to track boundaries of polygon structures with fixed shape and variable size through fields of pixel gradients and corners detected in images. Each match combines many attempts, based on sequential Monte Carlo sampling, to track a single polygon boundary. A polygon match is produced whenever at least one pair of adjacent corners is detected. The particle state variables capture geometric information about polygon matches, and the measurement variables capture photometric information derived from pixel gray values in the image. Measurement emphasis can be adjusted so that match rankings are based more on evidence of corners or more on evidence of edges in the image. We enable boundaries to be tracked efficiently by showing that for n-sided polygons of fixed shape, the number of required edge similarity calculations per polygon match can be reduced to only n no matter how many sequential Monte Carlo attempts are made to track its boundary. Several compelling match results against diverse clutter backgrounds are provided.

Important Links:

Go Back