Motion data retrieval from very large motion databases

The reuse of motion capture data has become an important way to generate realistic motions. Retrieval of similar motion segments from large motion datasets accordingly serves as a fundamental problem for data-based motion processing methods. The retrieval task is difficult due to the spatio-temporal variances existing in human motion. With the increasing amount of data, the retrieval task has become even more time consuming. In this paper, we present a motion retrieval approach that is capable of extracting similar motion subsequences from very large motion databases given a query motion input. Our method employs BIRCH-based(Balanced Iterative Reducing and Clustering using Hierarchies) clustering method to incrementally cluster poses so as to effectively deal with very large datasets. An elastic LCS(longest common subsequence) algorithm is then proposed to discover the similar motion subsequences based on the posture clustering result. Finally, the motion patterns are extracted and stored, with each pattern containing a set of similar motions. In the runtime retrieval stage, as each stored pattern effectively compared with the query motion, the group of the similar motions is acquired. Experimental results show that our method successfully retrieves similar motions and outperforms the existing methods in time and space costs when applying to very large motion datasets.