Object shape recognition approach for sparse point clouds from tactile exploration

Abstract

In this paper a novel approach is proposed for tactile shape recognition, which uses tactile point location and normal information. Superquadric functions are applied to construct several shape primitives and k-means unsupervised clustering method is used to partition the objects as several patches. By extracting geometrical features from each patch and rearranging features, object feature vectors are constructed for Gaussian process (GP) classifier to identify object shapes. Simulations results prove that our approach can achieve a high recognition rate in object shape classification task from sparse and noisy tactile point clouds.