Similarity Retrieval of Motion Capture Data Based on Derivative Features

Abstract

In this paper, we propose (1) a method of similarity retrieval of motion capture data in which a new feature extraction technique is introduced for the improvement of similarity search precision, as well as (2) a method to reduce the search time on a large database by using lower bound Dynamic Time Wrapping (DTW). For similarity search, joint speed has been mainly used as features of a particular motion. Our method differs from others in that we use not only the magnitude of speed but also the pattern of speed change. We measure the pattern of changing joint speed in a short period of time with the derivative of joint speed. In our experiments, we found that our proposed feature extraction can improve search precision and time. The average precision was greater than 90% and its computation time was 10 seconds on a dataset of 225 motion clips with a total of 81,851 frames from CMU’s database. The experiments showed that we can improve search precision using our proposed feature extraction technique compared to the retrieval method without using this method. For search time, our experiment shows that our retrieval method using the lower bound DTWcan efficiently reduce the amount of search data.