Human Skeleton Extraction Based On Machine Learning

Methods on human skeleton extraction can be applied to many aspects,such as motion capture and motion recognition and analysis.Traditional methods on human skeleton extraction requires additional devices like optical devices or sensors which brings about high cost.In this thesis,human skeleton extraction technology refers to image-based human skeleton extraction based on machine learning,which means three-dimensional human skeleton can be extracted automatically from image.Existing methods on human skeleton extraction mainly focus on extracting two-dimensional human pose.And there is no research which estimating three-dimensional human pose using machine learning technology.In this thesis,a novel framework to extract 3D human skeleton from single image using convolutional neural network is proposed.The framework integrates the technology on human region extraction from image,architecture of convolutional neural network and post-processing bone optimization.Furthermore,a method for extracting human skeleton from a sequence of images is proposed,which focuses on optimizing human skeleton based on spatial-temporal continuity.Experiment results show that human region extraction methods in this thesis has a good ability of extracting human region in high accuracy.And the algorithm of human skeleton extraction can fast predict human 3D skeleton from a single image,meanwhile the results are almost the same with ground truth in some poses.