Research On The Prediction Of Human Motion Trajectory Based On Neural Network Toward Human-robot Coexistence

The industrial robots have been widely used in automated operations such as palletizing,sorting,processing and welding.However,due to its bulky structure and high deployment cost,it is difficult to adapt to frequent alternation of production lines due to product customization and diverzification.The automation solution using human-robot collaboration can effectively reduce the deployment cost of the production lines,release the duty of workers,and meet the growing demand for flexible manufacturing.To ensure the safety of humans and the efficiency of robots in the Human-Robot coexistence(HRC)environment,it is necessary to accurately predict the trajectory of humans.The existing "look-forward" method for prediction of human motion trajectory model updates the parameters of the semi-adaptive neural networks(NN)model based on the estimation of predicted trajectory error.However,as the information of past trajectory is not fully ulitlized,the error of trajectory prediction cannot be accurately accessed and the model parameters cannot be accurately updated either.This degrades the accuracy of trajectory prediction.In this thesis,a novel "look-back-and-look-forward" NN model method is proposed to predict the trajectory of people in the HRC system NN model.This method firstly estimates the model parameters in the past(k-M)-th time step based on the actual error of trajectory prediction in the past.Subsequently,the model parameters is recursively updated till the current k-th time step.So that the trajectory till(k+M)-th time step in the future is accurately estimated,as well as the mean square estimation error of trajectory.After development of above trajectory estimation method for single particle,we further extend this method to estimation of the trajectory of a rigid body's pose based on fitting of an ellipsoid's parameters.In order to verify the effectiveness of the proposed algorithm,the hypothetical planar trajectory of a single point and the motion data of human torso from CMU Graphics Lab are used for simlation.Finally,experiments on motion data collection and algorithm verification are conducted at Motion Capture Lab,in Ningbo Institute of Materials Technology and Engineering,Chinese Academy of Sciences.This also provides guidance for setting the safety set of robot operations in the HRC environment and designing robot control algorithms.