Research On Human Action Recognition Method Based On Kinect

With the development of modern computer science and technology,the demand for human-computer interaction is getting higher and higher.Human action recognition is an important way to realize human-computer interaction.Its development has also become a hot issue in society today.Many researchers make a lot of research on this subject.At present,human action recognition has broad prospects for development in intelligent monitoring,medical treatment,games,and education.With the birth of Kinect technology,human action recognition will also be promoted to another level.The researchers study human action recognition based on grayscale and color images at early time.However,because they are susceptible to noise such as light,color,texture,and shading,which can reduce the recognition accuracy.Therefore,it is a new target for people to find a new type of human movement characteristics.The advent of the Kinect depth sensor brings us deep images with both human skeleton and depth information.This article studies the method of human motion recognition based on the joint point data extracted by Kinect.This paper reads as follows:(1)It summarizes the background and significance of the subject of human action recognition.At the same time,it briefly expounds the main process of human body motion recognition:human detection,feature extraction and motion recognition algorithms.Finally,the application of human action recognition based on Kinect is introduced.(2)According to the effect of convolutional neural network model in the large-scale image recognition competition,how to transform the action feature information into an image as its input.This paper proposes to simplify the action-space and time-space information into an image with four tracks,which called the distance change timing diagram.Each track from left to right to reflect the size of a distance variation with time,and finally using the traditional convolutional neural network to extract distinctive feature from it for human action recognition and classify actions.(3)In order to deal with the variability within the same actions and the similarity between different actions,an improved Citation-KNN action recognition method is proposed,which mainly adds the joint importance information and timing information of the action to the Hausdorff distance so as to increase the tolerance to noise and time migration.Before that,the dimensional catastrophe of the Citation-KNN algorithm is also alleviated by extracting the key frames in the action sequence.(4)The two methods are studied on the Berkeley MHAD dataset and the UTD-MHAD dataset respectively.The advantages and disadvantages of the two methods are compared and analyze the reasons.