Efficient unsupervised estimation of second-order B-spline contour descriptors

A novel unsupervised algorithm is presented for efficient and global computation of periodic second-order B-spline approximations to closed boundaries. Key local geometric information is extracted from a smoothed version of the boundary. This local information allows intelligent partitioning of the boundary and construction of an initial system of equations that often produces a very good approximation. Additional equations are introduced as local constraints to control occasional violations of the user-specified absolute error tolerance. The overdetermined systems of equations are solved by a standard least-squares approach. Computational complexity is compared to two previous algorithms. Experimental results are also provided.