A study on the decomposition of human motions into movement primitives, and application on gesture recognition of robot Poppy using Kinect motion capture

Beta Process Auto Regressive Hidden Markov Model

stage été 2016 avec PhamMQ

backup data and matlab code in backup1:stage_HMM_MovementDecomposition_Mpham

In my project, i study on general framework of segmentation algorithms, then focus on one method which is most appropriate to the objective of project that is to decompose physical exercises into movement primitives so robot can learn and reproduce exercises easily. I use Beta Process Auto Regressive Hidden Markov Model to model the generation of human motions from movement primitives. It is a Bayesian non parametric approach for problem of modeling time series. Then I use Markov Chain Monte Carlo to infer heuristic proposal segmentation from data. The number of primitives, the set of primitives and segmentation will be discovered without any information about movements. Finally, I classify and generalize learned primitives so that robot can determine whether recently detected primitive is learned and ask user for labeling if it is unknown primitive. After identifying primitives, robot will translate those data into dynamical movement primitives in order to be able to reproduce those primitives.

Not so good, problem of timescale.

completely simulated data on kinect : PMQ was moving arms and legs in different orders.