Research Of Key Frame Extraction And Posture Transition Technology On Mocap Data

In recent years, as comput er technology develops ra pidly, mocap(motion capture) tec hnology is widel y use d in game animation, film and television production, sports training, medical re habilitation, culture protection and so on. With the development of mocap technology, schola rs do further study of motion capture technology no matter at home or abroad gradually. This paper studies ke y frame extraction and posture transition te chnology in depth and the resea r ch result s a re mainly mani fested in two as pects:1. Existing key frame extraction methods often fail to re veal the local topological structure of motion c apture data. To this effect, thi s paper present s a LS(Laplacia n score) based on feature selection approach to e xtrac t the key frames from the moti on capture data. The bas ic idea is using the ideas of LS feature selec tion to select discriminant charac teri stics of human movement and finally get more genera l key frames. The proposed approach first extracts and grades two kinds of representative and normalized feature vect ors from t he original mocap da ta, and thus without distur bance of abundant features, it generates a comprehensive characteristic curve from the more important features, wit h the concave a nd convex poi nts on the curve corresponding to t he candida te key frames. With the c onstraint of the time threshold and the strategies to discriminate simi lar pose s, we further utilize the improved k-means algorithm to cluster the candidate key frames so that the final key frames can be obtained except redundant ones. The experime ntal r esults have shown that the key frames extra cted by proposed approach are typical, and can make up the lack of the same interval s ampl ing method, and bet ter avoid the interfer ence from the characteristics of the simpl ified method of layered cur ve. It can be well summarized and suited to reflect the whol e mot ion capture data visua lly.2. Existing posture transi tion methods often ignore the motion states. To this effect, this paper proposes a Hidden markov model(HMM) c ombine d with Bayes rule approach i n post ure transition of mocap data. It regar ds tha t human behaviorscan be divided into several or dered states. And it will find a suitable transition based on the rule of the state s equence. This approach first gets the assembl ed feature matrix by extracting two kinds of representative and normalizing feature vectors from the ori ginal motion sequenc es which can represent the mocap data briefly. The n,it forec asts the hi dden elements of the assemble d feature matrix by the HMM model and finds the hidde n state s’ r ule. Afte r that, it sel ects the perfect motion segment with the Bayes rul e based on the forecast result for linked segment and finally makes some interpolation to achieve further smoothing. The experiment results have shown that the posture transition by proposed approach can keep the motion sequence state rule well, generate a natural post ure t ransiti on, and avoid the slid, c ross and penetration movement distort ion, with t he junction having no obvious trac e.This paper has done simulation experiments with the moca p data from public dataset CMU and HDM05, and hence proved proposed methods’ effect iveness by comparing wi th other methods.