Kinect Depth Data Segmentation Based On Gaussian Mixture Model Clustering

Depth image data directly reflects the3D geometric information of the surface ofthe objects in the scene, and is not affected by light, shadows and other factors. Thisprecise spatial geometry information makes it easier to create a model of some object,simplify the algorithm and the processing method, and therefore be more conducive tothe development of the various disciplines of depth image. The acquisition technologyof depth data is varied, although these obtained data have high quality, the system costis higher. Compared with the traditional technique, the Kinect sensor released byMicrosoft is not only small, portable, fast reaction and cheap, but also the quality ofdata can meet the needs of the scientific research and application. All of theseadvantages make the Kinect attract widespread attention, improve the research workabout the Kinect depth data, and computer vision system based on depth image isincreasingly causing for concern.Indoor scene understanding based on the depth image data is a cutting-edge issuein the field of three-dimensional computer vision. Simplifying the structure of thediscrete three-dimensional data, extracting high-quality image features, are the basisfor the realization of three-dimensional object recognition and positioning.Segmentation is the most important step before feature extraction, and plays animportant role in the data preprocessing step, this work also can lay a good foundationfor the latter parts of the object recognition.Taking the layout characteristics of the3D indoor scenes and more plane featuresin these scenes into account, this paper presents a depth image segmentation methodbased on Gaussian Mixture Model clustering, aimed at extracting planesapproximated by the data of the scene. First, transform the Kinect depth image datainto point cloud which is in the form of discrete three-dimensional point data, denoiseand down-sample the point cloud data; second, calculate the point normal of all pointsin the entire point cloud, then cluster the entire normal using Gaussian Mixture Model,and finally implement the entire point clouds segmentation by Random SamplingConsensus algorithm. Experimental results show that the divided regions haveobvious boundaries and segmentation quality is above normal, and the sets of planes in3D space obtained from the previous operations lay a good foundation forfollowing object recognition.