Mining Motifs from Human Motion

Abstract

Mining frequently occurring temporal motion patterns (motion motifs) is important for understanding, organizing and retrieving motion data. However, without any a priori knowledge of the motifs, such as their lengths, contents, locations and total number, it remains a challenging problem due to the enormous computational cost involved in analyzing huge motion databases. Moreover, since the same motion motif can exhibit different temporal and spatial variations, it prevents directly applying existing data mining methods to motion data. In this paper, we propose an efficient motif discovery method which can handle both spatial and temporal variations of motion data. We translate the motif discovery problem into finding continuous paths in a matching trellis, where each continuous path corresponds to an instance of a motif. A tree-growing method is introduced to search for the continuous paths constrained by a branching factor, and to accommodate intra-pattern variations of motifs. By using locality-sensitive hashing (LSH) to find the approximate matches and build the trellis, the overall complexity of our algorithm is only sub-quadratic to the size of the database, and is of linear memory cost. Experimental results on a data set of 32,260 frames show that our method can effectively discover meaningful motion motifs regardless of their spatial and temporal variations.