Signal Classification in Quotient Spaces via Globally Optimal Variational Calculus

A ubiquitous problem in pattern recognition is that of matching an observed time-evolving pattern (or signal) to a gold standard in order to recognize or characterize the meaning of a dynamic phenomenon. Examples include matching sequences of images in two videos, matching audio signals in speech recognition, or matching framed trajectories in robot action recognition. This paper shows that all of these problems can be aided by reparameterizing the temporal dependence of each signal individually to a universal standard timescale that allows pointwise comparison at each instance of time. Given two sequences, each with N timesteps, the complexity of the algorithm has a cost of O(N), which is an improvement on the most common method for matching two signals, i.e., dynamic time warping. The core of the approach presented here is that the universal standard timescale is found by solving a variational calculus problem related to differential geometry.