Research Of Human Pose Estimation Based On Convolutional Neural Network

Human pose estimation refers to the detection of key parts and main joints of the human body in the image.It is the key to the human action recognition and behavior analysis,and it has broad application prospects in computer vision and other directions.The human pose estimation technology is more and more mature,but it also faces many problems.The images with poor lighting conditions or complicated backgrounds have different degrees of similarity between the appearance of the human body and the background;under different lighting conditions,different human appearance and the different movement postures make the same part of the human body show different appearances;the occlusion of various parts of the human body and the mutual occlusion between the human body and the objects cause the appearance of the human body to be incomplete;if there is no relevant prior knowledge,various parts of the human body may be exists in any area of the image,the search detection area is larger.Relying on the depth camera to quickly capture human gestures,the recognition and processing of these gesture actions can realize human-computer interaction,but it depends on specific sensors,the application scene is limited,and usually human-computer interaction based on cheap cameras is more desirable.This thesis takes the monotonous camera image human pose estimation as the main research content.Firstly,according to the structural characteristics o f human pose,the representation model of human pose is given.The concept and principle of Pictorial Structures Model and deep learning are studied.The application and advantages of the two methods in human pose estimation are analyzed,and the theoretical basis of human pose estimation is formed.Secondly,construct a convolutional neural network structure,define middle-level detectors in conjunction with contextual features,and learn them to impr ove basic positioning.Use these detectors to extract spatial and score-related features that are assumed relative to other body parts.The Set Boost classifiers recalculate the probability score for a particular part to determine its location.Finally,the commonly used human pose estimation datasets are summar ized,and experimental data are selected to perform human pose estimation tests to verify the feasibility and stability of the method.From two aspects to analyst the human pose estimation results,the accuracy of each part of the body and the accuracy of each joint.Design and analyze comparative experiments with other pose estimation methods.The experimental results show that the human pose estimation method studied in this thesis has human body parts and joint detection capabilities.The results of multiple sets of experimental tests for different test sets have certain stability.The accuracy of human pose estimation is explained by two different standards for each part's metrics and for each joint's metrics.Compared with other human pose estimation methods,the accuracy has been improved.