Cloud point labelling in optical motion capture systems

Abstract

This Thesis deals with the task of point labeling involved in the overall workflow of Optical Motion Capture Systems. Human motion capture by optical sensors produces at each frame snapshots of the motion as a cloud of points that need to be labeled in order to carry out ensuing motion analysis. The problem of labeling is tackled as a classification problem, using machine learning techniques as AdaBoost or Genetic Search to train a set of weak classifiers, gathered in turn in an ensemble of partial solvers. The result is used to feed an online algorithm able to provide a marker labeling at a target detection accuracy at a reduced computational cost. On the other hand, in contrast to other approaches the use of misleading temporal correlations has been discarded, strengthening the process against failure due to occasional labeling errors. The effectiveness of the approach is demonstrated on a real dataset obtained from the measurement of gait motion of persons, for which the ground truth labeling has been verified manually. In addition to the above, a broad sight regarding the field of Motion Capture and its optical branch is provided to the reader: description, composition, state of the art and related work. Shall it serve as suitable framework to highlight the importance and ease the understanding of the point labeling.