Imitation of human body poses and hand gestures using a particle based fluidics method

Download
2012
Tilki, Umut
In this thesis, a new approach is developed, avoiding the correspondence problem caused by the difference in embodiment between imitator and demonstrator in imitation learning. In our work, the imitator is a fluidic system of dynamics totally different than the imitatee, which is a human performing hand gestures and human body postures. The fluidic system is composed of fluid particles, which are used for the discretization of the problem domain. In this work, we demonstrate the fluidics formation control so as to imitate by observation initially given human body poses and hand gestures. Our fluidic formation control is based on setting suitable parameters of Smoothed Particle Hydrodynamics (SPH), which is a particle based Lagrangian method, according to imitation learning. In the controller part, we developed three approaches: In the first one, we used Artificial Neural Networks (ANN) for training of the input-output pairs on the fluidic imitation system. We extracted shape based feature vectors for human hand gestures as inputs of the system and for output we took the fluid dynamics parameters. In the second approach, we employed the Principal Component Analysis (PCA) method for human hand gesture and human body pose classification and imitation. Lastly, we developed a region based controller which assigns the fluid parameters according to the human body poses and hand gestures. In this controller, our algorithm determines the best fitting ellipses on human body regions and human hand finger positions and maps ellipse parameters to the fluid parameters. The fluid parameters adjusted by the fluidics imitation controller are body force (f), density, stiffness coefficient and velocity of particles (V) so as to lead formations of fluidic swarms to human body poses and hand gestures.