A Method for Measuring Three-Dimensional Human Joint Movements in Walking

A method was developed for converting a stereo-pair two-dimensional human key-points recognized by OpenPose to three-dimensional ones. Severe noise included in them due to fluctuation in OpenPose recognition and/or missing values were surprisingly reduced by four procedures; missing value interpolation, low rank filter, puppet model, and spatial AR model. The method gave us enough performance to analyze personal character in walking.