Master-Thesis-Motion-Recognition-in-Human-Robot-Interaction-using-Invariant-Descriptors
The motion of a rigid object is expressed in a way, in which itwill hold certain invariant properties with respect to several contextual dependenciesof the recorded motion. The primary goal of the thesis is to investigate and validatethe invariant properties of the motions through the recognition approaches that areproposed, followed by the juxtaposition and ranking of all the possible invariantrepresentation schemes, that are investigated, in terms of recognition accuracy. Atthe same time, an equally important objective, is to prove that the higher the levelof invariance a motion representation has, the less training motion demonstrationsare required in order for this motion to get recognized.The motion recognition algorithms that are proposed, consist of a modified ma-chine learning, distance-based, and a deep learning method. The amount of motionsrequired to train the corresponding methods is kept minimal and equal for everyapproach, so that a comparative basis can be established, and the invariance of theinvestigated motion representations can be discussed.
