Research On Keyframe Extraction And Retrieval For Motion Capture Data

With the popularity of Motion Capture System since 1990s, motion capture (MoCap) data can be gained directly, which results in the construction of massive motion databases. So, there is requirement for keyframe extraction and retrieval methods for MoCap data to compress raw data efficiently and use and administrate those motion databases.The work of this thesis is to explore new approaches to extract keyframes from MoCap data and retrieve MoCap data and presents the following algorithms (methods), including bone angles as motion features, layered curve simplification based keyframe extraction method from MoCap data, and contend based retrieval of MoCap data. The thesis is organized as follows:In the first chapter, we introduce the motivation, the state of the art and the brief description of this work.In the second chapter, we make a brief survey on motion capture technologies and keyframe extraction and retrieval methods of Mocap data.In the third chapter, firstly we introduce the motion description model, and then we give a brief survey on current motion feature representations. At last, we present bone angles as human motion features.In the fourth chapter, we present a layered curve simplification based keyframe extraction method from MoCap data. At first, we find keyframe candidates based on bone angles. After then, a new layered curve simplification algorithm will be used to refine those candidates and the keyframe collection can be gained. In addition, we employ an adaptive algorithm parameter to meet different requirements for compression.In the fifth chapter, we present a framework for retrieval of MoCap data. The framework comprises of two parts: preprocess of MoCap database and motion matching. In the first part, keyframes are extracted from raw MoCap data and storedin the MoCap databas, and then, index maps for every limb will be built. In motion matching, by index maps some postures similar to the start and end frames of query will be searched. Then those similar postures will be put in timeline and be filtered and combined to motion clip candidates. At last, elastic matching will be applied between those candidates and query to get final retrieval results.In the sixth chapter, we conclude the work and discuss future work.