On the Generalized Frechet Distance and its Applications

Measuring the similarity of spatio-temporal trajectories in a sensible fashion is an important building block for applications such as trajectory clustering or movement pattern analysis. However, typically employed similarity measures only take the spatial components of the trajectory into account, or are complicated combinations of different measures. In this paper we introduce the so called Generalized Frechet distance, which extends the well-known Frechet distance. For two polygonal curves of length n and m in d-dimensional space, the Generalized Frechet distance enables an individual weighting of each dimension on the similarity value by using a convex function. This allows to integrate arbitrary data dimensions as e.g. temporal information in an elegant,flexible and application-aware manner. We study the Generalized Frechet Distance for both the discrete and the continuous version of the problem, prove useful properties, and present efficient algorithms to compute the decision and optimization problem. In particular, we prove that for d is an element of O(1) the asymptotic running times of the optimization problem for the continuous version are O(nm log(nm)) under realistic assumptions, and O(nm) for the discrete version for arbitrary weight functions. Therefore the theoretical running times match those of the classical Frechet distance. In our experimental evaluation, we demonstrate the usefulness of the Generalized Frechet distance and study the practical behaviour of our algorithms. On sets of real-world trajectories, we confirm that the weighting of the spatial and temporal dimensions heavily impacts the relative similarity, and hence the ability to tailor the measure to the application is a useful tool.