Relationship Network Based 2D Human Pose Regression Algorithm

Human pose estimation is an important basis for image analysis and behavior recognition The detection of the position of human joints in the image is used for the subsequent auxiliary understanding of the image content.With the continuous intelligentization of electronic devices,more and more human behavior analysis is used in the fields of intelligent monitoring,human-computer interaction and motion analysis.Therefore,it is of great significance to study it.In view of the problems of occlusion and large degrees of freedom of joints in human pose estimation,the current 2D human pose estimation algorithm does not take into account the mutual connection between human joints,but directly extracts global features through the convolutional neural network and then goes to predict the position of each joint point.This leads to the loss of the inherent connection between the joint points.This thesis believes that the information between joints should receive more attention.Existing regression network algorithms do not make use of the constraint information between the joint points when matching.When the regression network extracts features,it is more from high resolution to low resolution and then upsampling to high resolution network model.The obtained feature map has a loss in accuracy.Based on the above analysis,this thesis gives a 2D human pose regression algorithm based on relationship network,the main work is as follows:1)2D human pose estimation based on local pose constraintsThe posture of the human body is affected by human movements.The freedom of each joint is large,and the search space between the joint points is large,resulting in the characteristics of ever-changing joint states.This thesis presents a 2D human pose estimation algorithm based on local pose constraints.The posture of the human body seems to be a whole process,but when it comes to a certain part,we can segment it from the whole and analyze it.The parts formed by the adjacent joints can be regarded as a small part of the human body.According to the model of the human figure structure,the local pose constraint relationship between the joint points is constructed.By constructing the local pose constraint relationship,more message transmission paths are added,so that the joint points can use more constraint information when matching.It was verified on the MPII dataset and achieved good results,with PCKh reaching 88.7%.2)2D human pose estimation based on relationship networkFor visible joints,local pose constraints can better constrain adjacent joints by learning structural information.However,there is the influence of occlusion interference between human joint points,and it is difficult to determine the positions of other joint points through local joint points.For this problem,this thesis presents a 2D human pose estimation algorithm based on relational network.Construct the relationship network of the human body through the similarity of the features between the joint points,and use their appearance features and geometric features to interact.The appearance features are expressed as the color features of the object itself.These geometric features represent the position of the object.There are multiple relationship modules,which can be compared to each layer in the neural network.There are many different channels in order to learn different types of features.The final output of the module is the fusion of the current joint points and the appearance features of each joint point.Each relationship module uses the two features of all joint points as inputs to obtain the combination of different relationship features and merge with the original feature information of the joint points.Experiments were carried out on the MPII dataset,and the joint points in the occlusion situation were significantly improved,of which PCKh reached 89.1%.3)2D human pose estimation based on multi-resolution fusionIn order to make the position of the joint point of the regression map of the feature map output by the network model more accurate,this thesis aims at the problem that the current regression network model is inaccurate to restore the position of the joint point of the regression of the high-resolution feature map from the low-resolution image generated from the high-to-low resolution network.A 2D human pose estimation algorithm based on multi-resolution fusion is given.In this thesis,the structure of the high resolution network is improved,and the multi-resolution structure is added to the model of the regression network.The multi-resolution structure uses high resolution to accurately locate the position of the joint point and low resolution to obtain more semantics.Based on the characteristics of the information,different resolution layers are fused to obtain the final output of the network.In order to verify the performance of the 2D human pose regression algorithm based on the relational network,this thesis conducted experiments on the MPII dataset.Experimental results show that the algorithm in this thesis successfully and effectively uses the information between the joint points of the human body to better realize the 2D human pose estimation.PCKh reaches 89.2%,which is better than the results of the comparison algorithm,indicating the effectiveness of the algorithm.