Research On 3D Human Body Estimation Based On Deep Learning

3D human pose estimation based on deep learning is an important research field of visual three-dimensional reconstruction.It is widely used in the fields of visual monitoring,sports rehabilitation and e-commerce,mainly including two-dimensional(2D)and three-dimensional(3D)human pose estimation.In the two research tasks,the accuracy of the former directly determines the accuracy of the latter,and the algorithm performance of the latter directly affects the visual effects and subsequent processing.The existing solutions require software to model and render the input image to complete.Although the visual effect is good,it does not solve the real-time tracking movement requirements,which is costly and consumes human resources;in addition,it can be directly combined from the video through the existing model The extracted data is used for modeling,although real-time requirements are solved,but factors such as human joint flexibility are not considered.Although the use of deep learning methods solves the above problems,a large amount of extremely difficult to collect 3D character annotation data is still needed to train the network.In order to solve the above problems,this paper introduces the prior knowledge of human motion mechanics,designs a function term with human body parameter constraints,improves the Generative Adversarial Network(GAN),and proposes an information fusion model to perform 3D human pose Real-time estimation,using back-projection technology can get a higher accuracy rate with a small amount of training data.The proposed method is not affected by the camera angle and the small amount of training data,and flexibly handles the complex human motion posture,and optimizes the network parameters.The main work of this paper is as follows:First,the 2D pose estimation method is used to extract the two-dimensional joint point coordinates from the image or its sequence,as the input of the GAN with the parameterized constraint loss function of the human body,to generate a 3D tree-shaped human skeleton,which reduces the pose ambiguity and errors in the network Parameters.In terms of network optimization,regularization is adopted to improve the generalization performance of the network,a local optimal solution is found to solve the mode collapse problem,and the Adam optimization algorithm is used to increase the calculation speed.Secondly,the fusion spatio-temporal feature method is used to estimate the pose of the moving 3D human body in real time,and the weighted average joint position error loss function is introduced to constrain the global motion trajectory to optimize the motion pose of the character.Back-project the processed 3D character skeleton onto the2 D plane,and compare it with the input coordinates to see if it is consistent,reducing the dependence on annotation data.Improve the algorithm's estimation accuracy of complex motion behavior in the video.Finally,in order to better train the model proposed in this article,I built a real environment data set with similar actions to the public data set.Using a variety of algorithms and the algorithm proposed in this paper are tested and analyzed in detail on the self-built data set and Human3.6 data set.The results verify the robustness and effectiveness of the method proposed in this paper.