Kinect Depth Map Inpainting Algorithm Based On Clustering Segmentation Of Color Image

In recent years,depth information has been widely used in target recognition like three-dimensional reconstruction,human-computer interaction and other computer vision systems.Microsoft Corp.'s Kinect is a low-cost depth sensor that can capture the depth map of the scene,and obtain a color image of the corresponding scene at the same time.More and more researchers are beginning to pay attention to Kinect.However,due to the limitations of its own imaging principle,Kinect obtains a depth image with depth information missing areas,a black hole area formed by occlusion,image optical noise.In order to obtain high-quality depth map and improve the application effect of Kinect,it is necessary to inpaint the depth map.In order to solve the problem of the inpainting of the Kinect depth map,this thesis combined the color image acquired synchronously and did the following research on the depth map inpainting algorithm,so as to improve the quality of the depth map and to inpaint the depth information of the real scene.This thesis first introduced the research background and significance of depth map inpainting,and the current research status of deep map inpainting algorithms,then analyzed some key issues in the depth inpainting method.The basic information of the Kinect sensor is introduced,the acquisition process of the Kinect depth map is analyzed,and the reason of the noise of the Kinect depth image is analyzed.The inpainting principles of several traditional inpainting algorithms are analyzed,and their inpainting effects are tested through experiments.There are a lot of holes in the Kinect depth map,but the traditional algorithm can not effectively use the global information and the neighborhood relationship provided by the color image,can not effectively segment the boundary of the object,produce a smooth phenomenon near the boundary of the object,and affects the inpainting effect of the depth map.In this thesis,considering the cluster characteristics of depth value,the color image was divided into multiple similar regions to obtain the structure information in the scene,so as to guide the depth map inpainting.The algorithm used C-means clustering to segment the color images acquired synchronously by Kinect to highlight the structure information of the objects in the target scene.The algorithm was based on the inpainting sequence of the fast marching algorithm and the void pixels were filled with pixels of the same type in the neighborhood of the void pixels.The effective pixel value is continuously diffused inside the cavity so that the entire hollow area is filled.Experiments show that the proposed inpainting algorithm has advantages over other algorithms in object edge strength and structural information accuracy.The quantitative indicators also show the effectiveness of the algorithm.Considering the fuzziness of the image and the ambiguity of the class to which the pixel belongs,a new segmentation of the color image was made basing on fuzzy Cmeans clustering.The idea of fuzzy clustering is introduced into the inpainting process and the fuzzy C-means is used.Clustering,regional segmentation of color images to obtain the membership function of each pixel,the segmentation results as a guiding image to guide the depth map inpainting.For the discrete void points in the depth map,the algorithm inpainted them by the bilateral filter,the brightness similarity factor in the depth estimation formula is changed to the membership degree factor.For the largearea void area,the structure information provided by the color image is introduced,and the weight function of the fast marching algorithm is redesigned.Experiments show that the algorithm has more advantages in edge retention,not only maintains the original target object structure information,but also more natural in the transition of the boundary,can clearly separate the foreground and background areas.