3D human skeleton extraction for action analysis

Abstract

This thesis is “3D Human Skeleton Extraction for Action Analysis”. We want to generate an accurate 3D pose estimation from a single view RGB image to understand more about people in images and videos. The 3D human pose estimation algorithm contained two steps. First, we put all of the 3D Human samples into K-means algorithm, and take the mean pose in each cluster as anchor pose ground truth in training part. Second, we trained first model called “3D anchor pose estimator”, which use 2D human pose as input ground truth, and the result of clustering as output ground truth. Anchor poses are some common human posture in our daily life. Final we train second model called “3D human pose estimator” and combine 2D human pose and 3D anchor pose to estimate final 3D human pose. In this work, we show a systematic design for how Fully Convolutional Network (FCNs) can be incorporated for the task of pose estimation. Our deep learning network contains two-stage: The first network (FCN1) estimate a 3D Anchor Pose from the 2D skeleton, then pass the results to the second stage (FCN2) to further regress/refine the 3D Anchor Pose to yield the final 3D human pose estimation. According to the experiments, our two-stage FCN network can generate a 3D human pose with an average MPJPE (Mean per Joint Position Error) of 62.99 mm when a 2D skeleton prediction is used as the input. The 2D skeleton predictions are produced by a pre-trained model called Stacked Hourglass. Stateof-the-art results are achieved on the H36M benchmark