Unsupervised Range Image Segmentation and Object Recognition Using Feature Proximity and Markov Random Field

In this paper, we propose a framework for unsupervised range image segmentation and object recognition that exploits feature similarity and proximity as leading criteria in the processing steps. Feature vectors are distinctive traits like color, texture and shape of the regions of the scene; proximity of similar features enforces classification and association decisions. Segmentation is performed by dividing the input point cloud into voxels, by extracting and clustering features from each voxel, and by refining such segmentation through Markov Random Field model. Candidate objects are selected from the resulting regions of interest and compared with the models contained in a dataset. Object recognition is performed by aligning the models with the refined point cloud clusters. Experiments show the consistency of the segmentation algorithm as well as the potential for recognition even when partial views of the object are available.