Research On Human Action Recognition Method Based On Kinect Skeletal Information

With the gradual development of computer intelligence,people's daily life has become more convenient and fast.Convenient life can not be separated from the support of various data.It is a way to study people's living habits to use computers to analyze people-centered data.People's living habits are mostly expressed by their behavior.Human action is a part of human behavior,so human action recognition is widely used in our real life.With the maturity of machine learning algorithms,human action recognition has become a hot issue in the field of human-computer interaction.In this dissertation,Kinect skeletal data is used to study different human actions.Firstly,a method of human action recognition based on two-dimensional planar projection features is proposed.According to the idea of three-view projection in engineering,this method uses three-view directional projection of human joints to find the characteristics of human motion.Analyzing human actions from different two-dimensional planes and multiple perspectives is helpful to solve the problems of human occlusion and single perspective defects.According to the law of human motion,joint vectors are constructed to express the motion of body parts.The angle between each joint vectors represents the relative motion relationship between the two body parts.The motion system of the whole body frame is represented by a combination of seventeen joint angles in three projection planes.Considering the time-varying relationship of each frame action,the joint angle information at each time is calculated by cosine similarity.Finally,according to the size of samples and the characteristics of motion characteristics,the appropriate classifier is selected to achieve the purpose of human motion classification.Based on the above research methods,a hierarchical feature fusion method for human action recognition is proposed in this dissertation.This method adopts the strategy of hierarchical recognition of human motion,divides human body into five parts from the biological point of view,and uses the combination of these five parts to represent human actions.Firstly,all the actions are roughly classified,and the motion features of the roughly classified are also based on the two-dimensional plane projection information of joint angle,so that some of the actions with great differences can be separated.Secondly,the human action is subdivided into several categories,which take into account the essence of motion and the relationship between the whole and the part of the human body,and the coordination between multiple degrees of freedom of human joints.The physical features of angular velocity and acceleration in kinematics are used to describe human motion.These features are fused from two different dimensions,which fully play a key role in the expression of motion details.Finally,two different classifiers are designed according to the attributes of motion features,which bring the better classification effect of the first layer to the second layer,thus increasing the stability of action recognition.Using the hidden information in human skeletal data,human actions can be accurately identified.Experiments on MSRAction3 D open dataset demonstrate the real-time performance and reliability of the proposed method,and show its unique advantages in identifying sequential actions.