Research On Head Pose Estimation System Based On Unmarked Motion Capture

The head poses estimation is a process of estimating the six degrees of freedom pose information of the key nodes of the human head in the process of movement.Its estimated indicators include the three-axis coordinate position and three-axis attitude angle of the key nodes of the head.The accuracy and robustness of head pose estimation,as well as the endurance of the head pose estimation system,will affect experiments and applications based on this technology,such as human motion research,rehabilitation therapy,and the use of augmented reality devices.Therefore,improving the performance of head pose estimation is an important issue in this direction.In recent years,there have been many methods applied to the field of head pose estimation,among which positioning and mapping-based pose estimation methods are more widely used.However,the performance of positioning and mapping methods heavily relies on device computing power and battery life,which limits their application.In practical applications,significant motion can also affect the loop detection trajectory correction module of positioning and mapping methods,increasing system error.At the same time,the optical active motion capture system is difficult to be widely used because of its high cost and need to be installed in a fixed area.To make up for the shortcomings of the above systems,the main work of this paper is to design and build a head pose estimation system based on unmarked motion capture and inertial measurement module.The system hardware consists of four Kinect Azure DK scene cameras and a low-cost inertial measurement unit.The software part consists of front-end and back-end systems.The front-end system uses the OpenPose algorithm to obtain human bone nodes from multiple scene cameras and obtains the three-axis coordinate position information of key human nodes through multi-frame iterative triangulation and a four-dimensional correlation graph algorithm.The back-end system uses the loose coupling method based on the error state Kalman filter and the inertial measurement module to obtain the optimal estimated six degrees of freedom position and attitude information of the head.Furthermore,to verify the accuracy and robustness of the system’s pose estimation,two experiments were designed in this paper:(1)Hand position accuracy experiment,which verifies its position accuracy by comparing the position information collected in the front end of this paper with the true value results of the OptiTrack system.(2)The head six degrees of freedom pose accuracy experiment,that is,the system combining the front and rear ends of this paper and two methods based on positioning and mapping are used to estimate the pose,and at the same time,the pose estimation accuracy is calculated with the truth value collected by OptiTrack system.Finally,the accuracy results of the three methods are compared.The experimental results show that in a complex scene such as a large six degrees of freedom movement of the head,the system proposed in this paper compares two methods based on positioning and mapping,and achieves an accuracy of less than 5cm under a stable state,while not producing trajectory drift and jump.Finally,the backend of this system only requires inertial measurement module data,without the need for additional high sampling rate cameras,thus having the advantage of long endurance.Compared to the OptiTrack system,it can significantly improve portability(smaller size,no need to install fixed equipment)and reduce costs while meeting accuracy requirements.