Research On Indoor Scene Segmentation Technology Based On RGB-D Data

Image segmentation intends to divide the image into several non-overlapping regions which have similarity features in the same region and significant differences in different regions based on the gray,color,texture and geometry characteristics of the image information.Image segmentation is a key research area in image processing.It is also the basis of pattern recognition and semantic segmentation and widely used in intelligent mobile robots for object grabbing and motion avoidance.Because of the complexity of scene in natural image and the subjectivity of human visual perception,different people have different understanding of image scene,so image segmentation has been a difficult problem in computer vision.With the popularity of 3D cameras in recent years,it is more and more easy to obtain RGB-D data,which makes the indoor scene segmentation based on RGB-D data get more and more attention.There are many technical problems that need to be broken down for RGB-D data segmentation in the form of Kinect point cloud.For example,the 3D point cloud data is large and the segmentation process takes a long time.In addition,due to the complexity of the 3D indoor scene itself,So the segmentation result is easy to produce flaws.In this paper,we focus on the above problems in the segmentation of RGB-D data,and study the algorithm based on super-voxel fusion and subsequent optimization.The main work can be divided into the following aspects:First,the supervoxel pre-segmentation is performed on the point cloud image.By setting the voxel size,the point cloud is transformed into voxel cloud,based on voxel,the voxel color,spatial coordinates and normal vector are extracted,and the voxel clustering result is obtained by clustering the voxels with similar characteristics.Since the super voxels are clustering of point clouds with similar characteristics and spatial adjacency,the subsequent point cloud segmentation is not based on a single point cloud as a processing unit,reducing the complexity of subsequent processing.Secondly,the supervoxels are clustered.In this paper,we use the extended convexity criterion to determine the concavity and convexity of the boundary between adjacent supervoxels.Based on the concavity and convexity of the connecting edges between adjacent supervoxels,the adjacent supervoxels are connected.Supervoxel clustering.Finally,according to some shortcomings in the results of the supervoxel clustering segmentation,an optimization scheme is proposed.The energy function composed of data loss,smoothness loss and tag loss is constructed by graph theory.By calculating the energy function Value to optimize the segmentation results,get the final fine segmentation results.The comparison of the experimental results shows that the segmentation method is better than other algorithms,and the accuracy of segmentation is improved,and the time consumption is small.In addition,this algorithm belongs to the unsupervised segmentation algorithm,compared with the supervised segmentation algorithm,without manual manual marking and sample training.