We construct human posture probability density based on actual human motion measurement. Human postures in daily life were measured for two days by having subjects wear a mechanical motion capture device. Accumulated human postures were converted to unit quaternions to guarantee the uniqueness of posture representation. To represent probability density effectively, we propose eigenpostures for posture compression and use the kernel-based reduced set density estimator (RSDE) to reduce the number of posture samples and construction of posture probability density. Before compression, unit quaternions were converted to Euclidean space by logarithmic mapping. After conversion, postures were compressed in Euclidean space. Applying constructed human posture probability density for unlikely posture detection and motion segmentation, we verified its effectiveness for many different applications.