The Study Of Real-time Motion Recognition Algorithms In Virtual Reality Systems

Real-time human motion recognition is an important and challengingresearch topic in the area of human-machine interaction. In immersive virtualreality systems, to make a more natural and effective communication betweenhuman and machine, systems need to capture and recognize human motions,and then parse them to syntactically correct complete motion commands.Nevertheless, most human machine interaction systems based on motionrecognition cannot offer satisfactory user experiences, with the lowperformances of motion capture and recognition in effectiveness and accuracybeing the bottleneck.This thesis focuses on the study of real-time3D human motionrecognition in immersive virtual reality games, and proposes several methodsto improve the original real-time motion recognition framework inrecognition sensitivity, accuracy as well as scalability during featureextraction and recognition process. The main work and contribution of thisthesis includes:First, a hierarchical motion training and recognition frame work isproposed, motions are divided into three different levels–L1M (the primitivemotion), L2M (the lexical motion) and L3M (the syntactical motion). This willbe more effective and accuracy for the overall recognition process, moreover,it offers the tolerance of spatio-temporal variations due to the insensitivity ofthe first class motion data to the aliasing and rhythm variation of motions, asit is relatively short in time and data. The original method, however, treatedall motions the same, this will cause trouble when a new motion class whichis quite different from the sampling window in length, because then DTWalgorithm will lose application value and may damage the motion data.Second, abstraction and feature extraction of motion data expose the major feature of a motion class, singular value decomposition has the powerof extract the principal geometrical characteristics which is a strong anddistinguished representation of each different motion class, so it is muchapplicable to motion data feature extraction, as opposed to the originalmethod, SVD will eliminate the much redundant data from a motion matrix,so this will offer more spatio-temporal tolerance.Third, this thesis proposes a dual-threshold screen based on within classand between class distances, together with SVM they form a combinedclassification method of motion recognition, this combined method has thehigh classification accuracy of SVM and the low computational complexity,thus offers high real-time performance and scalability to the framework.Fourth, a stochastic context-free grammar based method is introduced toanalyze the motion class streams, it can not only provide useful feedback toguide the recognition process, but also help understand the motion classes assyntactically correct and meaningful motion commands.At last, we testify the framework proposed in this paper with series ofexperiments and data analysis, the results show that, the framework andmethod within can provide a real-time and effective recognition of humanmotion.