Motion Capture Data Based Human Behavior Segmentation And Motion Retrieval

In recent years, with the continuous development of the motion capture technology and the wide application of motion capture system, people have accumulated a large amount of motion capture data, which has been widely used in animation, computer games, film production, and other industries. Due to the costly expense and complex operation, so how to effectively manage and reuse the existing motion capture data has received wide attention of researchers both at home and abroad.This thesis focuses on studying human behavior segmentation and motion retrieval based on motion capture data, and provides corresponding algorithm for every aspect of the problem. These algorithms have been validated by experiments. The main work of this paper includes the following three sections:(1) For the problem of human behavior segmentation, a new behavior segmentation algorithm based on cosine distance is presented. Angle histogram is used to describe the statistical characteristics of the human motion capture data, and cosine distance is utilized to measure the similarity of motions represented by angle histograms. First, for a given human motion sequence, a sliding window which moved from the starting frame to the end is defined. In each window, the cosine similarity between the first half window and the second half is computed, and so a cosine distance curve with the motion is obtained. The minimums of the curve identify the cut between different types of motion behaviors. The experiments show that our behavior segmentation algorithm can achieve96%precision rate and93%recall rate.(2) For the problem of motion retrieval, a new motion template synthesis algorithm based on angle histogram is proposed, which is used for motion retrieval through matching with the unknown motion sequence. First, given several motion sequences of same type and calculate the angle histogram for each motion sequence. Since the angle histogram of all the motion sequences have the same dimension, so directly calculating the average of these angle histograms can obtain the motion template of corresponding motion type. The key property of motion template is that it grasps the essence of motion sequences of the same type, so it could be used for motion retrieval. The experiments show that the average precision rate of motion retrieval based on motion template8%higher than motion retrieval based on dynamic time warping distance.(3) We developed a behavior segmentation system with C++graphical user interface application according to our cosine distance based behavior segmentation algorithm. Through the demand analysis, this system can be divided into data processing module and interface display module. The data processing module realizes the motion capture data format conversion, segmentation point detection and motion sequence segmentation. The interface display module displays the result of behavior segmentation and provides interactive method to adjust the segmentation points. This behavior segmentation system could improve the accuracy and efficiency of behavior segmentation of motion capture data greatly.