Robust Object Recognition Based on Regular Framing and Depth Aspect Image

A method of model-based object recognition for a cluttered depth scene including multiple objects is proposed. A novel model representation, named depth aspect image, is also defined as an orientation standardized appearance from the original depth data of objects, which is transformed through tuples of three barycenters defined within regularly defined voxels. A robust matching scheme, named least quantile of residuals, can achieve not only object recognition with depth aspect images but also verification with candidate models. The sparsely distributed barycenters and the robust matching make the ICP-based rough registration and the following verification process much faster and more reliable. In this paper, we show recognition experiments on 100 scenes contained multiple objects from a library of 4 models.