Hypercube sweeping algorithm for subsequence motion matching in large motion databases

Current optical motion capture devices are capable of capturing motion at frequencies exceeding 1000Hz thereby generating gigabytes of motion data. In this paper we propose a method to solve the problem of subsequence motion matching in large motion databases. Our method supports non-uniform time-scaling. We begin with a polar-angle representation of the motion that gives a continuous thread in a multi-dimensional space. We improve the performance of the matching process by generating a motion curve index based on a representation of multiple 1-D signals rather than by partitioning the multi-dimensional space into subspaces as done in some previous work. Given a motion query, we sweep a hypercube along the query thread. Motion subsequences intersected by the hypercube form a matching set. Our method matches any possible non-uniform time-scaled subsequences between the query and the database, since any non-uniform time-scaled motion retains the same shape and location of the thread in the multi-dimensional space. We propose a new method to perform fast hypercube sweeping by utilizing a histogram. The histogram counts how many dimensions of each point on the thread are matched. A point is inside the hypercube when its count equals the total dimension d. The histogram is incrementally updated to minimize the sweeping cost. Our results show that the performance of our method depends on the speed of the query motion. We stress test our method by streaming the query motion against a motion database to determine its performance. The results show that the system can handle larger databases on slower query motion and vice versa.