Modeling of human motion through motion captured data using non-linear identification

Abstract

The importance of estimating human motion analysis can be illustrated by numerous applications such as performance measurement for human factors engineering, posture and gait analysis for training athletes and physically challenged persons, animation of the human body, hands and face, automatic annotation of human activities in video databases, control in video games and virtual reality or teleoperation of anthropometric robots. Image processing technique from motion captured images is an accurate and cost effective method to give a set of data that defines the location of specified limb at every sequence of human motion. From this set of data, system identification was done to model the human motion. This project is a study on how performance of an identified model is influenced by different types of model representation whether it is a linear model or non-linear model and a single variable model or multi variable model. Two types of parameter estimator was used which were least square and recursive least square. The study also included the effects of different number of lags on the model. The objective is to formulate a predictive model to analyze human motion. Simulation studies were done on this model representation and compared with actual human motion. Several model validation techniques were done to validate the identified models. In this study, multivariable non-linear model is a good human motion representative. The model accuracy increases as the degree of non-linearity and number of lags are increased but it makes the model become more complex.