Recognition Of Moves From3D Human Motion Capture In Interactive System

In interactive systems, the most important is how human and computerinteract. The facts that influence the experience of human computerinteraction are whether the approach of interaction is nature, whether thereaction from the computer is correct and whether the response from thecomputer is real time. As a result, human motion recognition is consideredas a very important branch of human computer interaction. And it is being acutting-edge research. In an interactive system, motion recognitionsub-system should take two aspects into consideration. One is the accuracyof recognition and the other is the response efficiency. The motionrecognition needs trade-off between accuracy and efficiency.This thesis illustrates a high-level framework of the motion recognitionsystem including the training, real-time recognition and the optimization ofrecognition. The framework keeps the low coupling between training andrecognition.First, according to the agility of human’s limbs, we introduce amulti-layer splitting method for the human skeleton model based on which weextract features that can be adapt to different users of different height andbody for all motions. This thesis starts with the source of human motiondata and the extraction of features. With high-precision, high-frequency andhigh-credibility human motion data, we construct a multi-layer humanskeleton model which can show the changes of a motion directly and fromwhich it is easy to extract useful features.Second, we propose new feature extraction method based on motioncurves with three kinds of parameters including the quaternions of humanbody, the relative quaternions to joints’ parents and the relative quaternions to symmetric joints. There are several difficulties in finding high-performancefeatures because the features that descript the motions need to be adapt todifferent users with different height and different body. Motion curves withthose parameters above as features can solve those problems. At last, we useAdaboost method to training for strong classifiers with positive and negativedata.Third, we introduce an optimized Dynamic Time Warping to prevent thebias of rhythm when calculating similarity between input motion and thestandard motion. It is more useful than the standard DTW because it hasconstraints on the offset of time.Fourth, we propose a top-down feature weight descending pyramidmatching for reducing the complexity of calculation during feature extractingand distance calculation because brutally calculating similarities of allfeatures once with Dynamic Time Warping algorithm is time consuming.Fifth, we propose a Hidden Markov Model based most likely motionclass tagging method which works as a smoothing filter added in the end ofthe recognition’s pipeline. Based on the candidate motion of previousmethod’s output, this optimization adds correlation between each recognitionoutput which previously is independent with each other and its aim is toreduce the overall noise (result with low credibility) of the outputs with thehelp of Hidden Markov Model.At last, we testify the effectiveness and efficiency of each method byexperiments and data analyses.