Abstract:
Three-dimensional (3-D) local feature description is an important task in 3-D computer vision. However, accurate and robust feature description is still a challenge in the presence of various nuisances (including noise, occlusion and clutter) contained in real scenes. To improve the performance of feature description, this paper proposes a novel LSVSH local feature descriptor. First, a new attribute (called curvature attribute) with independent of LRA is designed, improving the robustness of LSVSH to the error of LRA. Second, the local space is divided into several bins along the radial direction, and then three angle attributes and one curvature attribute are counted in each bin to generate the LSVSH descriptor, achieving a full and robust description of local surface information. Extensive experiments on B3R, U3M, U3OR and QuLD datasets are implemented. Results show that the AUC
pr values of LSVSH on B3R, U3M, U3OR and QuLD datasets are 0.95, 0.70, 0.54 and 0.10 respectively, which are all superior to existing local feature descriptors.; LSVSH achieves 70% correct registration on U3M dataset and 100% correct recognition on U3OR dataset, verifying its effectiveness in these applications.