Articulated Pose Estimation Based On Spatial Transformation Convolution Neural Network

Nowdays,due to the technical requirements of people tracking and motion recognition,attitude estimation has become one of the most important computer vision challenges.At present,researchers have proposed a variety of human pose estimation algorithms.However,due to difficulties such as light changes,human body deformation,parts occlusion,high dimension of the state space and large size,human pose estimation needs to be further studied.In this dissertation,a method of human joint attitude estimation based on spatial transformation convolution neural network is proposed.The spatial transformation convolution neural network model is constructed to train the joint dependency and improve the classification accuracy.The main works are as follows:Firstly,the data pretreatment.In this dissertation,the data set of LSP(Leeds Sports Poses)is selected.In order to obtain a larger data set and to improve the generalization ability of the model,the pictures in the dataset are rotated and turned over to enlarge the size of the data set.Then the images are divided by the center of the joints to get the joint image pieces,which are clustered.The clustering results are used as labels for the subsequent deep spatial transformation convolution neural networks to train the joint dependencies.In the experiment,K-MEANS algorithm and K-MEANS ++ algorithm are used to do the clustering comparison.The experimental results show that the K-MEANS algorithm is simple and the result is better for two-dimensional joint image blocksSecondly,the deep spatial transformation convolution neural network is used to train the joint dependencies.In this dissertation,a deep spatial transformation convolution neural network is constructed based on caffe framework of deep learning.Then K-MEANS clustering results are used as the input of the network to train the joint dependencies.In the training through cross-validation of the way to constantly optimize the network parameters.The experimental results show that the convolution neural network with deep space transform has better fitting ability to data,better joint dependence of training.The training of joint dependence accuracy increased by 10%.Finally,a decision-making function of Support Vector Machine(SVM)is built for joint's classification.This dissertation analyzes the function of decision-making function in SVM,and then constructs a score function based on K-MEANS clustering results and deep spatial transform convolution neural network training results through mathematical modeling.The score function is used as the decision function of SVM for joint classification.The score function is composed of two parts,one part has a preliminary judgment of a certain joint fragment according to the extracted picture features and the other part further identifies the fragment according to the joint dependency of training.The joint point is identified by a two-part overlay.The experimental simulation shows the accuracy of the prediction of the joint part and the joint point under different evaluation criteria.The average detection accuracy of the limb part of the PCP standard is increased from 75.0% to 79.8%.Under the PDJ@0.50 standard,the average detection accuracy increased by 0.5% to 89.0%.