Research On Human Action Recognition Based On Half-Body Mixed Model

With the advent of the Internet of Things and the Internet + era,people gradually into the era of intelligent network,which also means that computers and other intelligent devices will be more concerned about the movement of the human body and interact with people.Human motion recognition has always been a hot research issue in the field of computer vision and artificial intelligence,and has a very good application in intelligent monitoring and advanced human-computer interaction.This thesis mainly studies the recognition of the human motion based on the half body mixed model.In the research of action recognition,this paper uses the method of human body gesture sequence to model the human body.Human motion recognition is divided into three steps: image feature extraction,human pose estimation and action recognition.In this thesis,we study and design a set of solving algorithm to the problems of feature extraction efficiency,posture estimation accuracy and timing modeling of action recognition.The main research contents are as follows:(1)The main idea of traditional HOG feature extraction method and its existing problems are analyzed.A feature extraction algorithm based on GPU parallelization is designed and the feature extraction experiment is completed for the problem that the traditional HOG feature extraction method has large computational dimension and slow speed.(2)A human body attitude estimation algorithm based on the half body mixing model is designed.The human body half body mix model consists of a group of human body components,using a spring connection between components,using a fractional function to represent the component matching similarity and spring deformation overhead.From the model analysis,the model establishment and the model main line,the details of the realization of the attitude estimation are studied in detail,such as reasoning,distance conversion,message passing,backtracking and non-maximum suppression.(3)The algorithm of recognizing human body action based on hidden Markov model is designed.The design of the behavior recognition model,the expression and training of the model are studied in detail,and the Bayesian network model of the human posture and the hidden Markov model of the human posture sequence are designed.Finally,the analysis of the action recognition method is completed in different data sets and comparative experiments.From the theoretical analysis and the final experimental results,we can see that this paper solves the problem of feature extraction speed,attitude estimation accuracy and motion identification timing modeling in human motion recognition.