Unsupervised time-series clustering of distorted and asynchronous temporal patterns

Most time-series clustering methods, such as k-means or k-medoids, are initialized by prior knowledge about the number of classes or by a learning step. We propose an unsupervised clustering technique based on spatiotemporal mean-shift and optimal time series warping using dynamic time warping (DTW). Our main contribution consists in combining a spatiotemporal filtering technique, which gathers similar and synchronized temporal patterns in image sequences, with a clustering algorithm that applies a trajectory constraint on the DTW associations, thereby discriminating between similar time-series that are temporally shifted or warped. We assess the method's robustness on synthetic data, and demonstrate its versatility on brain magnetic resonance and multispectral satellite image sequences.