基于高斯混合模型聚类的 Kinect深度数据分割

摘要

Indoor scene understanding based on depth image is a cutting-edge issue in the field of three-dimensional computer vision.In 3D indoor scenes the planes are quite many, taking this feature into account, we present a Gauss mixture model clustering-based depth data segmentation method, and realise planes extraction from scene data.First, the method converts the depth image data acquired by Kinect into discrete three-dimensional data point cloud, and applies denoising and downsampling treatment on the point cloud data; On this basis, it calculates the normal vectors of all points in entire point cloud, and clusters the normal collection of entire 3D point cloud using Gaussian mixture model;next, it carries out the plane fitting on each clustering with random sampling consensus ( RANSAC) algorithm, gets a couple of planes from each clustering, and eventually segments the whole point cloud data into some sets of planes.Experimental results show that the divided regions using this method have accurate boundaries and the segmentation quality is above normal.The sets of planes extracted from the previous operations will lay a good foundation for the following indoor object recognition and scene understanding.%基于深度图像的室内场景理解是计算机视觉领域中的前沿问题。针对三维室内场景中平面较多的特性，提出一种基于高斯混合模型聚类的深度数据分割方法，实现对场景数据的平面提取。首先将Kinect获取的深度图像数据转换为离散三维数据点云，并对点云数据作去噪和采样处理；在此基础上计算所有点的法向量，利用高斯混合模型对整个三维点云的法向集合聚类，然后利用随机抽样一致性算法对各个聚类进行平面拟合，由每个聚类得到若干平面，最终把整个点云数据分割为一些平面的集合。实验结果表明，该方法得到的分割区域边界准确，分割质量较高。提取出的平面集合为以后的室内对象识别和场景理解工作奠定了较好的基础。