Research On 3D Human Pose Estimation Based On Deep Learning

Human pose estimation is one of the current research topics of computer vision.It uses the collected image information to detect the body parts and form a complete posture.With the change of computer technology,smart devices capable of generating image information are gradually widely used in daily life.As the main target of visual information,how to effectively identify the action and behavior of human is an urgent problem to be solved.According to the image feature extraction method of human pose,human pose estimation is divided into traditional methods and methods based on deep learning.According to the difference of human pose dimensions,human pose estimation tasks can be divided into two-dimensional tasks and three-dimensional tasks.Although various models and methods have been proposed to improve the detection accuracy,it is not enough to achieve the desired result,especially when the research task is applied in the actual scene,which is greatly restricted by conditions such as occlusion and environment changes.Therefore,it is an important task to improve the human posture estimation method.This thesis analyzes the shortcomings of the current human pose estimation method,and combines the methods of group normalization and multi-tasking to improve the existing dense three-dimensional human estimation model.The main contributions and innovations of this thesis are as follows:(1)Strategy of multi-scale fusion.The complexity of human body's structure,the diversity of movements and behavior,and the photographer's shooting angle all affect the scale of human target in the image.Using the high-level network to extract image detail information,and the global information extracted from the underlying network.By combining the two part of information,the model can effectively improve the accuracy of the estimation of the target human body and reduce the false detection and missed detection of the network.(2)Adoption of group normalization.Limited by hardware conditions,the size of Mini-Batch in batch normalization operation can only choose a smaller value,and it is not possible to obtain valid statistical information between different batches in the dataset,and the local mean and variance will cause serious errors to the final overall detection.In this thesis,the group normalization method is introduced into the three-dimensional dense human pose estimation task,which makes full use of the information of each channel of the image itself and improves the task detection accuracy.(3)Multitask cascading.Based on the idea of multi-task training and the characteristics of deep learning input data transformed into non-linear and abstract feature representation after passing through these hidden layers,this thesis analyzes the feasibility of multi task learning in deep learning research,and proposes a region based network structure in which multi tasks are executed simultaneously and their outputs are input to each other,so as to improve the accuracy of DensePose human pose estimation task.