We propose a parameter-free method for the detection of dominant points and polygonal representation of possibly noisy curves. Based on \cite{ND07,NguyenDebled09}, this work aims at a parameter-free method through a multiscale approach. We propose a new evaluation criterion to automatically determine the most appropriate width parameter for each input curve. Thanks to a recent result \cite{FF08} on the decomposition of a curve into a sequence of maximal blurred segments, the complexity of this algorithm is $O(n\log n)$.