Recognition of 3-D objects from multiple views
Henry Stark, Ron Wu · Annual Meeting Optical Society of America · 1985
Given an unknown test object and several reference classes, how do we determine which class the unknown object belongs to? The problem is more complicated when the object is 3-D and its size can vary over a broad range. Our approach is based on principles of tomography: Several views of the object are taken and reduced to projections (line integrals). The projections are Fourier transformed and mapped into a scale and rotation-invariant function via the Mellin transform and circular harmonic coefficients. The resulting function has other invariants useful in pattern recognition. A demonstration using two look-alike gears of different sizes is furnished. The technique is also useful if only partial view data of the unknown object are available. This is demonstrated as well.