Research on 3D Human Pose Estimation Using RGBD Camera

To aim at the problem of many researchers have only focused on recovering 3D human body information from color images, which is not accurate, causing great ambiguity and slow. we propose a new method for 3D human pose estimation. We get color images and depth images through RGBD camera. we use convolutional neural networks for 2D human pose estimation to get joint points coordinates in color image and then map the returned results to corresponding depth image to obtain 3D joint points information. For 2D human pose estimation, we improve the accuracy of the stacked hourglass network using Faster-RCNN and residual structure Resnet50 as the human target extractor. During the mapping process, a sparse feature point matching method based on the SURF algorithm is used to determine the calibration parameters of color images and depth images.