Accelerating Matching and Learning of Eigenspace method

We propose a method for accelerating the matching and learning processes of the eigenspace method for rotation invariant template matching (RITM). To achieve efficient matching using eigenimages, it is necessary to learn 2D-Fourier transform of eigenimages before matching. Little attentions has been paid to speeding up the learning process, which is important for applications in which a template changes frame by frame. We propose two key ideas: First, to further speedup the matching process using FFT, we decompose rotated templates to orthogonal fast-eigenimages using Fourier basis by utilizing the circularity of rotated templates. Second, to speedup the learning process, we compute 2D-Fourier transform of the fast-eigenimages in polar coordinates using Hankel transform[11]. Proposed learning method is equivalent to but considerably faster than that existing method, i.e., rotated template generation, SVD and 2D-FFTs in Cartesian coordinates. Experiments revealed that the learning, matching and the total processes becomes respectively 120, 3, and 36 times faster while keeping comparable detection rate compared to existing method utilizing SVD in Cartesian coordinates. The algorithm was successfully applied to global localization of mobile robot where online learning is required.