Spatio-Temporal Convolution Kernels for Clustering Trajectories

We propose a novel class of kernels to identify tactical patterns in multi-trajectory data such as soccer games. Formally, we introduce a group of R-convolution kernels called Spatio-Temporal Convolution Kernels composed of a temporal and a spatial kernel. The particular choice of the component kernels depends on the application at hand. For the purpose of clustering player and ball trajectories in soccer we propose a probability product kernel on the empirical distributions of the objects to serve as spatial kernel and a Gaussian kernel as temporal kernel. Empirically, we observe better clusterings compared to baseline methods and high cluster consistencies with (inefficient) Dynamic Time Warpingbased methods. In terms of tactical patterns we identify interpretable clusters corresponding to long and short game initiations on either sides of the field.