Action Recognition And Real-time Interaction Technology Based On Skeleton Information

Human computer interaction refers to the process of information exchange between people and computers.Nowadays,human-computer interaction plays an important role in the era of artificial intelligence,and it is developing towards user-centered,personalized biometrics and all-round perception.With the development of computer technology,human-computer interaction has played an important role in Internet products.In various interaction modes,natural human-computer interaction emphasizes human-centered interaction,such as voice interaction,body language interaction and other gesture language interaction,which makes the communication between human and computer more natural and smooth,so the application prospect is more visible.Natural human-computer interaction includes body sense interaction,which is a kind of computer equipment to "perceive" people's various behaviors,languages and expressions,so as to realize the exchange of information between people and computers.In this paper,based on the premise of somatosensory interaction,by recognizing the static action of the hand,and then taking the moment of recognizing the action of the hand as the beginning of the interaction,tracking the continuous movement track of the hand,so as to realize simple real-time interaction.As a branch of artificial intelligence,human motion recognition is widely used in many fields,such as human-computer interaction,virtual reality,abnormal behavior detection,intelligent home and motion analysis.Therefore,more and more scholars will put their time and experience into the research.Among them,there are some problems in the natural human-computer interaction based on human posture,such as low accuracy of motion recognition,poor differentiation of continuous motion and so on.So this paper proposes a method of human motion recognition and real-time interaction based on skeleton information.This method first obtains the two-dimensional coordinate information of 18 key points of each action,and then uses the stack model training data to complete the recognition of human motion.Finally,according to the recognized hand action,it completes the track tracking,there are seven real-time interaction schemesbetween human and machine,including four kinds of one hand interaction and three kinds of two hands interaction.Compared with video motion recognition,motion recognition based on skeleton information can effectively improve the recognition ability,and solve the influence of traditional methods,such as illumination,camera angle change,occlusion and motion interference.It has strong practicability and plays a decisive role in the future augmented reality.In this paper,the motion recognition method is divided into three steps.Firstly,the key points of human body are detected.In this paper,18 key points of human body are detected by convolution neural network of VGG and two branches,and the real two-dimensional coordinate information of key points is obtained.Then the movement recognition is carried out,the data of different types of movements are collected to make datasets,and the stacking model is built for training.In order to increase the complexity of experimental data,when collecting data sets,people of different shapes are selected as the collection objects.Due to the complexity and time-consuming of a large number of twodimensional coordinate training,in order to reduce the adverse effects caused by special sample data and make the data neat and consistent,it is necessary to normalize the data.Finally,in the process of motion recognition,we also track the specific movements,such as the hand movement track,and in the method of motion recognition,compared with the LSTM method,we find that the method used in this paper can recognize the movements with high accuracy.Through motion tracking,the motion trajectory is estimated,and the interaction instructions that the computer can adapt are fitted,so as to achieve the purpose of real-time interaction.