Research On Key Problems Of Action Imitation Of Humanoid Robot Based On Kinect

With the continuous development of intelligent robot technology,people are constantly expecting a better interactive experience.Human-computer interaction based on somatosensory has played an important role in improving intelligence and comfort.And the human action imitation is one of the research focus in this field.In this paper,we use the NAO_V5 robot of Aldebaran Robotics as the experimental platform,and use the KinectV2 depth camera as the somatosensory acquisition device to design the interactive system.And we conduct in-depth research on data acquisition,motion mapping and hand motion recognition control in the imitation process.This paper includes the following four parts.Firstly,the depth image and bone data of the human body are acquired by the Kinect.The position data of the joint is recorded.These data is transmitted to the NAO robot through the coordinate space conversion to realize the primary motion simulation under the human-computer interaction.Secondly,aiming at the jitter and tracking error problems in data acquisition,an improved nonlinear constrained Kalman filter algorithm is proposed.The algorithm incorporates the nonlinear bone length constraint into the tracking method of Kalman filter,and solves the problem of stable convergence of parameters by the maximum likelihood estimation method.Experiments show that the mean square error of the estimated state decreased from 6.06 dB to 5.77 dB,and the performance evaluation of the standard deviation of bone length change increased from 23.1% to 94.31%.So the algorithm effectively improves the smoothness and anti-interference of the data.Again,the article optimizes the imitation action algorithm.For the limitation of the physical structure of NAO robots,the inverse kinematics algorithm based on Newton iteration method is adopted.By comparing the difference between the target position and the current position obtained by the forward kinematics,the difference is minimized under the constraints of kinematics and dynamic conditions.The final mapping angle satisfies the motion similarity and its physical limitation.The effectiveness of the optimization algorithm is verified by an experimental evaluation based on the imitation of typical actions.Last but not least,for the problem of poor acquisition due to hand position limitation in real-time imitation,the hand movements at different positions are identified by extracting the histogram of gradient of the image and training the classification model with the support vector machine.Experiments have shown that the results of offline testing are 92%,and real-time online testing also shows higher recognition speed and accuracy.