Object Segmentation Based On RGBD Information Of Logistics Scene

Object segmentation is a foundation issue for robot sorting in logistics scenes.In traditional robot sorting jobs,objects are rigid with known CAD models or texture models.However,in the logistics sorting scenario,the logistics objects has the characteristics of unknown texture model,unknown shape,and overlapping.Therefore,it is difficult to complete the segmentation by using the traditional predefined CAD model matching algorithm or the texture matching algorithm.In this paper,the object segmentation algorithm based on RGBD information in the logistics scene is researched,which is based on single view object segmentation and pose estimation,multi-view fusion segmentation,and dynamic scene fusion segmentation.The research achievements of this thesis are as follows.1.An improved depth segmentation algorithm based on point cloud geometry feature and the surface feature is proposed.For objects with similar color and texture,there are some errors in segmentation based on color information.In this paper,we extract stable geometric information to complete segmentation.Compared with the traditional method of point cloud based on vector angle region growth algorithm,for one thing,based on the local feature of point cloud(mean point cloud and neighborhood point cloud angle),global features(point cloud normal vector and the point of view angle of line)and Euclidean distance similarity feature complete object surfaces segmentation.For another,the graph model is constructed by calculating the concave and convex relation and the connection relationship between the surfaces and completed object segmentation by merging surfaces.Experimental results show that the proposed algorithm achieves 91.3%accuracy on OSD dataset.2.A segmentation algorithm based on multi-view global segmentation model is proposed.In this paper,we use the RGBD SLAM framework to estimate the pose of the camera in multi-view.Because there are some errors in the estimation results,the same point in the real world has sin-gularities in the pixel positions in different perspectives.Therefore,this paper establishes a global segmentation model for storing pixel label values at different perspectives and completes label val-ue updating through multiple observations to complete the multi-view object segmentation in the voxel model.The experimental results show that the accuracy of multi-view segmentation is up to 96.84%compared with the single-view segmentation accuracy of 91.31%with the improvement of 5.53%.3.A fast segmentation algorithm based on the fusion of color,brightness,depth,background modeling,and foreground extraction is proposed.Traditional background modeling methods,such as GMM and codebook,which use the RGB information to complete the background modeling,are easily affected by light and shadow.Based on this,this paper proposes a background modeling that integrates color distortion,luminance distortion,and depth distortion algorithm.Because the logistics sorting scene often only changes locally,the global segmentation will be relatively time-consuming,so this article combines the segmentation result of the foreground region and the seg-mentation result of the background static region.The experimental results show that the algorithm has an average precision of 94.79%on the SMB data set and the image single-frame segmentation speed has risen from 1.27fps in a single view to 11.9fps.