Efficiency and accuracy tradeoffs in using projections for motion estimation

This paper presents an investigation of the use of the Radon (projection) transform in speeding up existing image registration techniques. The ultimate goal is to make these algorithms more computationally simple, while simultaneously realizing acceptably accurate performance. The use of the projections in estimating translational motion decomposes a 2D problem into a pair of 1D problems, leading to significant computational savings. Here we present the tradeoffs of computational efficiency and accuracy for two current methods. Our experiments show that for most applications, the modifications we suggest in using the projections instead of the image directly cost little in performance, yet realize dramatic improvements in computational efficiency.