Research On Action Understanding Method Based On 3D Human Pose Estimation

With the rapid development of computer technology and the improvement of people's demand for intelligent equipment,human action recognition has become one of the hot research directions in the field of computer vision.It is widely used in public security,human-computer interaction,virtual reality,sports,medical health and other fields.Therefore,human action recognition has a high theoretical research value.The early research work mainly focused on RGB video images,but it was difficult to achieve excellent results due to the influence of complex background and illumination intensity.However,with the development of depth sensor technology,it is convenient and feasible to obtain high precision 3D skeleton joint information.Compared with traditional RGB video image data,skeletal posture information has a great advantage in describing behavior.Skeleton node data can not only accurately describe human posture and motion state,but also is not affected by background complexity and illumination intensity,so skeleton information is widely used in action recognition.This paper studies the action recognition based on human skeleton nodes,and the main work is as follows:(1)This paper proposes to use human body structure and human joint motion characteristics to predict the missing node data.The normalized joint vector and the angle between joint vectors are extracted as the pose features,and finally the keyframes are selected by K-Means clustering algorithm.In action recognition,the selection of keyframes not only reduces the redundancy of data and the number of features in the process of action recognition,but also enhances the expression of the meaning of behavior.Hence,the selection of keyframes affects the accuracy of action classification,directly.(2)The keyframes selection is an optimization problem,this paper proposes to transform the keyframes selection problem into an optimization problem in binary coding space.In this paper,an evaluation model based on the fusion of domain information and the number of key frames is designed,which divides the behavior sequence into multiple domains.The model adaptively adjusts the number of keyframes according to the compression rate while keeping the motion timing.A Multi-Population based Multi-Objective Differential Evolution Algorithm(MMDE)is designed,and the concept of inflection frame is proposed.The inflection frames are used as the population initialization identification and the population initialization rules are redefined.At the same time,the differential mutation operator and the selection operator are improved to improve the global search ability of the algorithm.(3)In order to verify the effectiveness of the method,experiments are carried out on three common data sets: MSR-Action3 D,Utkinect-Action and Florence3D-Action.In the action recognition based on K-Means keyframes selection,the effect of action recognition on different classifiers is compared.Compared with the original sequence and the sampling frames of equal interval,the accuracy of action recognition is improved.Action recognition experiments of selecting keyframes based on optimization method,this paper compares the recognition effect with the mainstream methods in the current literature.The function of inflexion frame is analyzed,and the effect of adding inflexion frames and the significance of domain division are discussed.In the problem of action recognition,the keyframes selection model proposed in this paper can reduce the feature redundancy of behavior sequence to a certain extent,enhance the expression of physical meaning of behavior,and effectively improve the effect of action recognition.