Partial Similarity Based Motion Capture Data Retrieval

In the field of motion capture data retrieval, most of the previous works focus on the utilization of global similarity. As a novel approach, partial similarity based motion capture data retrieval algorithm has gradually got more and more researchers’ attention.In this paper, we propose a novel framework of partial similarity based motion capture data retrieval. First, we build a motion capture database that contains a large number of human motion capture data. Then, for every segment of motion capture data in the motion capture database, we treat a set of adjacent data frames as a basic unit, according to which we construct a serious of data matrices. Every column of the data matrices represents the motion information contained in the corresponding set of adjacent frames. By using sparse coding and dictionary learning algorithms, we can estimate the motion dictionary and its corresponding coefficient matrix for every data matrix. Every column of the data matrices can be approximated by the linear combination of several columns in the corresponding motion dictionary, coefficients of the linear combination are recorded in the corresponding position of the coefficient matrix. Then, according to these motion matrices and coefficient matrices, we can construct a serious of reference matrices. Every row of a reference matrix record the information of reference pattern of the corresponding motion capture data in the database with respect to the corresponding motion dictionary. With these information, we can retrieval the database.In the query process, for a given query sample, we get its reference patterns with respect to every motion dictionaries by sparse coding technology. After that, we can calculate a score for every motion capture data in the database by evaluate the similarity between its reference pattern and the query sample’s reference pattern. Then, the result of retrieval is returned in descending order according to the scores. Because the proposed motion capture data retrieval algorithm relies on the reference patterns generated by a serious of motion dictionaries, the results are not related to the duration and the order of appearance of motions. Further, the proposed algorithm take different length of motion segments into consideration, which also increases the performance of the proposed algorithm. Experimental results show that the proposed algorithm has achieved good performance.