There are numerous approaches that solve the registration problem; from the extensive exploration of the range images' degrees of freedom to the most commonly used ICP. However ICP is only able to work correctly when the images are close enough to the correct position. Another problem is the pairwise nature of registration, meaning that each corresponding range image match must be known beforehand for a complete batch of scans. This work proposes a novel approach that is able to align a set of unordered range images without knowing the correspondence between them to the point that ICP is able to succeed in its optimization. Consisting of a multilevel hierarchical structure parallel to the ones used in collision detection problems this work aims to reduce unnecessary processes in registration pipeline by simplifying the point clouds and evaluating and filtering possible matches at the different levels of hierarchy as well as filtering possible matching points by comparing surface descriptors. The results found from the experiments conducted to the proposal verify both numerically and visually that the outcomes are effective with real and virtual range images.
展开▼
