Research On Human Skeleton Extraction And Sitting Activity Recognition Based On Kinect

In recent years,with the rapid development of contemporary computer science and technology,computer vision has become increasingly popular.In addition to basic color video,Kinect can also provide corresponding depth image and skeleton data information.This information has made significant progress in the study of human activity recognition.The depth image information can additionally reflect the distance information from the observation object to the camera center and ignore the color difference between the clothes and the background;it's also insensitive to changes in the light.In addition,the human skeleton formed by the joint definition of the Kinect through 20 joint points also follows this essence that the emergence of posture and activity is due to changes in the human skeleton.This paper analyzes and studies the human skeleton extraction based on depth data and the human sitting activity recognition based on skeleton information around the various data information provided by Kinect.The main research work of the thesis is as follows:(1)This paper proposes a multi-layer human skeleton extraction algorithm based on depth image data delamination.In view of the self-occlusion of the limbs faced in the skeleton extraction in the current Kinect system,we use the data features of 2.5D of the Kinect depth image and divide the image segmentation process into two steps.The first step is to apply the thresholding principle to obtain the depth threshold for the depth image data and then perform the initial segmentation.Then the second step uses the difference between the depth threshold and the human body image region limit depth value to calculate the new depth threshold,and then perform secondary depth image segmentation to optimize the segmentation results of the human body region depth image that is ultimately involved in skeleton extraction.The experimental results show that the proposed method can solve the problem about self-occlusion of human body to a certain extent,and obtain a relatively complete human skeleton with no redundant branches.(2)This paper presents a human sitting activity recognition method based on skeleton sequence features.In the framework of this method,we first simplify the definition of Kinect skeleton joints and extract the spatial and temporal descriptors of local features based on skeleton information,and then use K-means clustering algorithm and principal component analysis algorithm to obtain the vector of locally aggregated descriptors of the feature clustering result.Then we use a custom loss function to propose a two-stage metric learning process using the global stochastic gradient optimization algorithm to obtain the final feature transformation result of the discriminant information.Finally,based on the metric transformation,the nonparametric k-nearest neighbors classifier is used to classify the feature information.The experimental results verify the feasibility and accuracy of this method in the recognition of human posture activity,and have a good performance in the human activity recognition.