Research On Human Motion Recognition And Prediction For Human-Robot Collaboration

With the development of industrial automation technology,robots are widely used in industrial scenes.The advancement of robotics has prompted robots to move from constrained environments to shared environments with humans.In a shared environment,humans and robots can interact and collaborate in various ways.A complete human-robot collaboration(HRC)system can improve production efficiency,improve the quality of task completion,and reduce human workload.When human and robot can communicate and understand each other effectively,they can achieve close HRC and get the best performance.In order to achieve safe and efficient HRC,it is necessary to make accurate predictions of human motion,so as to provide useful information for robot motion planning in advance.In addition,this paper hopes not only to recognize and predict human motion,but also to estimate the uncertainty of human motion.Aiming at the accuracy of human motion prediction,the human motion recognition and prediction method for HRC is studied in this paper:(1)Research on the acquisition of human motion data and feature extraction.According to the hardware structure of Kinect depth vision sensor and the principle of skeletal tracking,Kinect sensor is used to track human body and obtain human motion information.Segmentation and feature extraction of human joint motion trajectories are carried out,and a human motion data set is further constructed.(2)Research on the recognition of human actions in HRC scenarios.In order to understand human intentions,it is necessary to predict the next human movement by identifying tasks performed by humans.In this paper,we use dynamic time warping method to recognize human actions.(3)Research on human motion prediction algorithms.In order for collaborative robots to perform motion planning in time and improve the safety and efficiency of HRC systems,it is necessary to predict the human motion trajectories.Therefore,a human motion prediction method that combines the optimized sliding window polynomial fitting and recursive least squares is proposed.This method reduces the prediction error by setting an appropriate window length and adding constraints on the increment of prediction value.In order to estimate the uncertainty of human motion prediction results,a data-driven human motion prediction algorithm is proposed.This algorithm combines representative trajectories with multi-step forward Gaussian process regression,which not only reduces the prediction error,but also estimates the uncertainty of the prediction result.(4)Conduct experiments on the recognition and prediction of human motion in HRC scenarios.The performance of the proposed algorithm for human motion recognition and prediction in typical HRC tasks is tested.Experiments show that the proposed algorithm is effective in HRC scenarios.