Author(s)

Xue Yang, Hongyang Yu, Wanjun Huang

Corresponding Author

Xue Yang

ABSTRACT

At present, the restoration of 3D human information based on a single picture or video mainly uses 3D human body reconstruction in view of deep learning, which has the problems of long calculation time and high hardware requirements. For ordinary monocular cameras, 3D human arm information recovery method based on image ranging and spatial geometry is proposed. First of all, MobileNetV2 neural network has the characteristics of lightweight and low latency. The MobileNetV2 is used for 2D human pose estimation, and the original activation function is modified to maintain a low amount of calculations and parameters compared to traditional networks. It also improves the recognition rate. Then, based on the principle of camera imaging and spatial geometry, the method of estimating the three-dimensional information from the two-dimensional information of the human body is studied to obtain the three-dimensional information of the human right arm. Next, the joint angle is calculated by the space vector method to realize the consistency mapping from arm to manipulator and determine the motion of the wrist joint. Finally, relevant experimental research based on the UR manipulator was carried out to complete the human-machine natural interaction experiment, which verified the feasibility and effectiveness of this scheme.

KEYWORDS

MobileNetV2; Person pose estimation; 3D human arm information recovery; Human-computer interaction; style; styling; insert