3D Human Motion Capture Based on Neural Network and Triangular Gaussian Point Cloud

In this paper, an optical unmarked motion capture method based on convolutional neural network and triangular gaussian point cloud is proposed to achieve accurate 3D human pose estimation. Firstly, the Direct Linear Transformation(DLT) method is used to calibrate the actual multi camera system and obtain the internal and external parameters of all cameras. Then the depth neural network Cascaded Pyramid Network(CPN) is used to extract the 2D human key points in the images collected by each camera in the system. Next the triangle positioning and the least square method are used to preliminarily obtain the 3D human key point coordinates, and then the 3D key points of human body are optimized by gauss point cloud method to get the accurate 3D results of human body.