抄録
3D pose recognition is important as the intelligent
function of autonomous robot. The pose recognition is achieved by two stages called an initial registration and a detailed registration. In the initial registration, a corresponding between the model and the scene is made roughly. On the other hand, in the detailed registration a highly accurate corresponding is made. In this paper two new invariant features called HSI_S (Horizontal Spin Image of Surface) and HSI_E (Horizontal Spin Image of Edge) are proposed, and parallel use of them is shown experimentally to improve generality of pose recognition. In addition to that, two different ICP algorithms called F-ICP and I-ICP are also proposed and parallel execution of these two ICP's is shown experimentally to improve robustness to clutter and occlusion. The reason of that is that since focus regions of F-ICP and I-ICP are mutually complementary, the parallel execution of them (hybrid method) leads to the accuracy improvement. In this paper, several non-hybrid methods are compared to the hybrid method through pose recognition (registration) experiment of objects like auto parts and bird models. As a result, the proposed hybrid method with HSI_S and HIS_E has been shown to be better than single method (non-hybrid method) in its performance. That is, the proposed hybrids method is dominant in both generality and robustness.
