The Research On Marker-based Human Pose Reconstruction

Human motion capture technology is one of the hot and difficult subject invirtual reality, computer vision and computer graphics. It is widely used in humananimation, motion analysis, human interactive, simulation training, biomechanics andso on. Human motion capture is mainly to catch the human motion, even human facialexpression. To make animation realistic, drive animation or virtual object model withthese motion or facial expression data directly. In this paper, we reconstruct humanpose in passive optical motion capture. The research including marker registration,missing markers predicting and the joint locations estimation, the main researchcontributions are as follows:At first, we propose a marker identification method for unlabelled motion data inpassive optical motion capture. Find correspondences between model data and firstmotion data to finish the process of registration of first motion data. Marker trackingis completed by spatial-temporal characteristics and local rigidity of markers on thesame segment limb. Because of human motion and noise in optical motion capture,the distances between markers may be changed, we do a correction on registeredmotion data. The experiments show that the proposed method can register motion dataeffectively.After then, we propose a method of missing markers prediction for missingmarkers in motion data. In this paper, according to the fact that small motion in twoconsecutive frames, we assume that motion parameters are constant betweenconsecutive frames on the same limb segment. We predict the missing markerpositions using visible markers on the same limb segment. In this method, we predictthe missing markers positions according to the first two motion data and the visiblemarkers in the current frame. The process is simple and fast, and it can be used inreal-time motion capture system.Finally, we present a method of joint locations estimation on calibrated markerdata. If three markers on one limb, while less than three markers on the adjacent limbsegment, we estimate the joint parameters using non-linear least square algorithm. Forthe reason of noise and experiment errors, the estimated joint locations may not besatisfy the constant length constraint. We correction the joint positions after the jointlocations estimation. Experiments show that the method can recovery human poseeffectively.