Motion indexing using coordination between essential actuators

We address the problem of organizing a partially annotated corpus of human movement by indexing and linking similar motion segments. The motion indexing problem consists in finding all pairs of similar motion segments in the corpus. Motion indexing is an important step towards full annotation of a motion corpus and, consequently, to the discovery of robust representations to humanoid motion. We propose an approach where each human action is associated with a set of essential actuators. The set of essential actuators may range from a set with one element (single actuator) to a set containing all actuators (whole body). The novelty here in our motion indexing approach is the use of coordination between actuators to discover similar motion segments and to find a minimal set of essential actuators for each segment. This approach respects the parallel aspect of human motion where actions are performed concurrently. This allows the combinatorial use of different actions at the same time. This approach considers only a minimal set of actuators to model each action which results in performance gains. In this paper, we present the design of a non-ambiguous similarity measure, two greedy heuristics, and an optimal algorithm for the construction of coordination regions that ultimately result in similar motion segments. The heuristics support the optimality of the proposed algorithm. We empirically validate the optimality and correctness of our algorithm in our experiments.