Research On 3D Human Body Reconstruction Algorithm Based On Monocular Camera

As a sub-field of computer vision,three-dimensional reconstruction of human body has developed rapidly in recent years,and its application has become more and more extensive.On the one hand,this work has provided a lot of help to all aspects of human daily lifes and work.On the other hand,in the field of scientific research,this work also provides a possibility for computers to understand the three-dimensional information of the human body and to recognize the three-dimensional world,which is the forefront of future computer vision changes.In order to solve the problems of long time-consuming and low accuracy of the existing three-dimensional reconstruction methods of human body,we propose a new dense 3D point cloud reconstruction algorithm of the human body using a monocular camera according to the existing depth information-based 3D point cloud reconstruction algorithm of human body.For the 3D human body reconstruction algorithm based on deep learning,we propose a new feature point extraction method,add new constraints,and further optimize the reconstruction effect of extreme poses.In addition,we also compare the advantages and disadvantages of those two reconstruction algorithms,and look forward to the future research work.The specific research content is as follows:1)Aiming at the problems of low accuracy and high time consuming of human body 3D point cloud reconstruction method and high dependence on equipment and light,we propose a method of 3D point cloud reconstruction of human body based on monocular camera.First,calibrate the monocular camera of the experimental equipment,and collect the human body picture sequence as input.Then select SIFT feature point descriptors to extract feature points and complete matching.To solve the problems caused by mismatch and monocular camera,preform image preprocessing such as background segmentation.Based on the reconstructed sparse point cloud,we propose a region growing point cloud densification method,and finally obtain a dense human 3D point cloud reconstruction.Finally,we analyze the advantages and disadvantages of human 3D point cloud algorithm based on monocular camera.2)We use deep neural network to extract human feature points for subsequent 3D model parameter estimation.As the basis of human body model reconstruction,two-dimensional feature point extraction is an important link.We first introduce the extraction of feature points of the face,and propose a method for matching feature points of multi-scale and multi-face pictures by combining cascade neural networks and lightweight networks.Then combine the cascade neural network and the hand joint point detection model to extract the hand joint points.Finally,we train a stacked hourglass network to recognize human joint points and foot feature points.For each part of the network,we perform the corresponding data set preprocessing,including scale transformation,image enhancement,training set fusion,etc.,and finally we complete the accurate detection of two-dimensional feature points.3)Based on the use of feature points to estimate the three-dimensional model of the human body,we propose new constraints to optimize the extreme pose reconstruction.First,the facial feature points,hand feature points and foot feature points that can reflect the reconstruction details are added to the 3D human body model pose restoration work to reconstruct a3 D human body model with correct posture and complete details.Then,for the extreme poses encountered in this subject,we propose some improved methods to optimize the effect of 3D reconstruction of the human body while ensuring efficient reconstruction.