Research On Problems In 3D Human Modeling

With the development of 3D data acquisition technologies such as 3D scanners and depth cameras,massive 3D models and related applications have attracted the extensive attention of researchers in computer graphics and computer vision.Among them,the processing and analysis of 3D human data is always the core issue.3D human models have been widely used in digital geometry,game production,film special effects and virtual reality.A large number of methods have been emerging in recent years regarding the application and research of human body,3D human reconstruction is of great significance.At present,the research on 3D human modeling mainly includes: human pose estimation and 3D human reconstruction from 2D images,3D human reconstruction from 3D human point cloud data,3D human shape and pose retrieval.For large human datasets,the human pose retrieval and recognition is important for human shape analysis and pedestrian recognition.In this paper,we introduce the 3D human reconstruction based on traditional methods and deep learning respectively,and proposes a 3D human pose retrieval method based on neural network.This paper proposes a new network structure,which first uses an encode network to extract the features of the input,then use the 3D human model parameters of the training data and the features to estimate the shape and pose parameters of the input human point cloud by a linear regression network.We use a parametric human model to generate the human meshes and obtain the training point clouds by sampling on the generated data.And use the parameter error and reconstruction error between the training data and the output to train the network.Finally,we use the pose parameters to 3D human pose retrieve.The experimental results show that our method exceeds the existing methods on a benchmark in several metrics.And our method can directly process point cloud instead of mesh,at the same time,we can reconstruct the input point cloud.