Research On Extraction Method Of Human Posture Skeleton Based On Point Cloud Data

At present,most applications based on somatosensory devices in the market are developed by using a single somatosensory device,and human-computer interaction is completed by recognizing human gestures and postures.Under a single device,the capture range is limited due to the limitation of the viewing angle of the device.At the same time,under a single device,the self-occlusion of the human body will lead to the problem that the human body posture cannot be accurately estimated.If multiple somatosensory devices are used to complete human posture estimation and motion recognition,the accuracy will be greatly improved.Skeleton extraction is an important step in posture recognition.Therefore,it is proposed to use multiple somatosensory devices to study human posture acquisition and skeleton extraction to make up for the problems of narrow viewing angle and self-occlusion under single device,and provide a good foundation for subsequent posture recognition.In this paper,multiple depth cameras are used to collect the point cloud of human body.The workflow is as follows: according to the characteristics of depth cameras,choose the appropriate number and location of cameras and configure the collection environment;Geometric calibration of multiple depth cameras by iterative nearest point algorithm using calibration objects;Multiple depth cameras are used to scan the pose of objects from multiple perspectives.Considering the high noise and incompleteness of point clouds caused by the limitations of depth cameras and acquisition environment,the depth images are preprocessed by filtering algorithm to achieve the purpose of data smoothing.After being converted into point cloud data,the point clouds are preprocessed such as denoising and downsampling.Then,the camera calibration information is used to register each part of point clouds to obtain rough registration results.Finally,the point clouds are globally optimized based on third-order supersymmetric feature matching to obtain the point cloud model of human body pose.ROSA algorithm is improved to solve the problem that the branches of human posture point cloud easily interfere with each other during skeleton extraction,which leads to skeleton dislocation and skeleton deviation from center due to subsequent processing.Firstly,the point cloud is reasonably segmented based on the idea of Markov random field,then ROSA points are calculated by using the method based on rotational symmetry axis for each segmented point cloud,and the skeleton obtained after smoothing and thinning is taken as the initial skeleton.The skeleton has high topological consistency with the human body point cloud model and can adapt to different postures.However,because smoothing and thinning operation will make the skeleton deviate from the central axis,an algorithm based on active contour model is adopted to optimize the coarse skeleton,and finally a skeleton with good centrality is obtained.