Prediction of Human Motion with Motion Optimization and Neural Networks

Abstract

Human motion prediction is a crucial foundation for efficient and safe human-robot cooperation. In order to improve the accuracy of long-term motion prediction, an approach combining initial trajectory prediction and motion optimization is proposed. The neural network model is applied to model the human hand motion for initial prediction. Then, we utilize the Gradient Descent method to optimize the motion considering the target set constraints. The motion's target point position is obtained from K-Nearest Neighbor (KNN) classification model. Experiments on real reaching motion data demonstrate that the proposed algorithm achieves a higher prediction accuracy compared with the traditional single neural network model.