Optimization of human tracking systems in virtual reality based on a neural network approach

The problem of determining the optimal number and location of tracking points on the human body to ensure the necessary accuracy of reconstruction of kinematic parameters of human movements in virtual space is considered. Optimization of the human tracking system in virtual reality has been performed to reduce the amount of transmitted information, computational load and cost of motion capture systems by reducing the number of physical sensors. The task of optimizing the number and location of tracking points on the human body necessary for the reconstruction of a virtual body model from a limited set of input points using numerical approximation of the regression function is set. An algorithm has been developed for collecting a large amount of data from a human body model in a virtual scene and from a motion capture suit in the real world. The smallest number of human body tracking points and their location were obtained using the proposed algorithm. Various neural network topologies have been trained and tested to approximate the regression relationship between a vector of tracking points limited in size (from 3 to 13) and a vector of 18 virtual points used for the complete reconstruction of the human body model. The necessary accuracy of reconstruction of kinematic parameters of human movements is provided at 5 and 7 input points. The proposed approach made it possible to use 5 or 7 physical sensors to build a model of the human body and restore the kinetic parameters of its movements in virtual reality. The approach can be applied to solving inverse kinematics problems in order to reduce the number of physical sensors placed on the surface of the object under study, to simplify the processing and transmission of information. By combining data from both the motion capture suit and the virtual avatar, the process of collecting information has been significantly accelerated, the volume of the training sample has been expanded and various patterns of user body movements have been modeled.