Human Pose Estimation Based On Convolutional Neural Network

Human pose estimation is a key step to understand human behavior in images and videos.It has vast application prospects in intelligent monitoring,human-computer interaction and other fields.The convolutional neural network has a great advantage over the traditional image processing methods in the field of machine vision.The human pose estimation model based on convolutional neural network has greatly improved the prediction accuracy of the traditional human pose estimation model,the Tompson model is the classic one.In this paper,an improved Tompson model is established for the problem of large kernel and non-full size joint heatmap in the Tompson model,and on this basis,the transfer learning is introduced to establish transfer learning human pose estimation model.The main work of this article is as follows:1.Based on the Tompson model,an improved Tompson human pose estimation model is established.The 3×3 convolutional kernel Tompson model is established by using the 3×3 convolutional layer group to take the place of 5×5 convolutional layer and 9×9 convolutional layer in the Tompson model.On the basis of the 3×3 convolutional kernel Tompson model,the full size Tompson model is established by using transposed convolution to output full size joint heatmap.2.The transfer learning VGG19-30 s and transfer learning VGG19-8s human pose estimation model are established.The feasibility of using VGG19 model to solve human pose estimation problem is analyzed.Based on the feature extraction part of the trained VGG19 model,a transfer learning VGG19-30 s model for human pose estimation is established.Based on the transfer learning VGG19-30 s model,the transfer learning VGG19-8s model is established by feature reuse.3.The mini-batch gradient descent method is used to train the models,and then test the trained model.The FLIC-plus dataset is selected as the training data of the model,and then using data augmentation and mean subtraction to preprocessing the data.According to the true coordinates of the human joint points,the true value joint heatmap is established by the 2D Gauss distribution.Design the cost function of the models.Then,the model is trained by the mini-batch gradient descent method.The PCK evaluation index is used to analyze the model prediction results.The validity of the improved method in this article is verified through the prediction accuracy comparison.