Research On Human Parsing Algorithm Based On Feature Fusion And Posture Correction

Human parsing is usually considered as one of the basic tasks of human-centered visual understanding activities or event recognition and detection.Nonetheless,a large gap seems to exist between what is needed by the real-life applications and what is achievable based on modern computer vision techniques.With the explosion of massive visual media data in daily life,human body parsing task has gradually become a research hotspot in the field of computer vision.This paper will focus on two different tasks in human parsing.The first part focuses on the human parsing algorithm in single instance visual scene;the second part focuses on the perception and parsing algorithm of multi-instance in complex visual scene.The key technologies of this paper mainly focus on multi-level feature fusion of human parsing and case perception and parsing based on human posture.The main research contents and contributions of this paper are as follows:In order to solve the problem of detail missing and semantic confusion in human parsing algorithm,a new double pyramid unit parsing network is proposed.Firstly,a spatial pyramid unit is introduced to capture the multi-scale information of two low-level features,so as to enhance the ability of the model to capture tiny targets and overcome the problem of detail missing in the process of human parsing.Then,a context pyramid unit is used to aggregate the local to global semantic information of two high-level features to eliminate the parsing confusion in the process of human parsing.The method achieves better performance than the existing methods on single person and multi person datasets.At the same time,the speed can reach 41.2 FPS.In the multi-instance scene,aiming at the problem of incomplete parsing caused by irregular human posture(such as back or tilt)and cross occlusion,a multi-person human parsing network based on human posture correction is introduced for the first time.Starting from the inherent characteristics of the human body,firstly,the key points of the human body are used to sense and segment multiple instances,and different instance regions are extracted.Then,the affine transformation is used to correct some human features with irregular posture,so as to obtain a complete and accurate multi person analytic reasoning result.The experimental results show that the average accuracy and the average cross union ratio of the proposed method are71.91% and 59.61% respectively.