Interpolated Rigid Map Neural Networks for Anatomical Joint Constraint Modelling

The demand for accurate individual and general kinematic joint models is increasing with growing applications in fields such as animation, biomechanics, motion capture, ergonomics and robot human interaction modelling. Many approaches have exploited unit quaternions to eliminate singularities when modelling orientations between limbs at a joint, leading to the development of a number of novel joint constraint validation and correction methods. A number of machine learning approaches have been applied to this modelling problem, as depending on training data either individual or general joint models can be created. Recent work has demonstrated the use of Rigid Maps to model regular conical constraints on the orientation of the limb. In this paper we extend this work deploying a modified Rigid Map Network with a continuous output.