دانلود رایگان مقاله اسپارس کارآمد ICP

Abstract

The registration of two geometric surfaces is typically addressed using variants of the Iterative Closest Point (ICP) algorithm. The Sparse ICP method formulates the problem using sparsity-inducing norms, significantly improving the resilience of the registration process to large amounts of noise and outliers, but introduces a significant performance degradation. In this paper we first identify the reasons for this performance degradation and propose a hybrid optimization system that combines a Simulated Annealing search along with the standard Sparse ICP, in order to solve the underlying optimization problem more efficiently. We also provide several insights on how to further improve the overall efficiency by using a combination of approximate distance queries, parallel execution and uniform subsampling. The resulting method provides cumulative performance gain of more than one order of magnitude, as demonstrated through the registration of partially overlapping scans with various degrees of noise and outliers.

8. Conclusions and future work

We have presented an efficient variation of the Sparse ICP algorithm that is based on a new hybrid optimization which starts with a general Simulated Annealing search and then switches to ADMM-based ICP, to ensure convergence to an optimal solution. We have also provided several insights on how to further improve the efficiency of our method using a combination of approximate distance queries, parallel execution and uniform subsampling. The cumulative performance increase over the original Sparse ICP is more than one order of magnitude when tested on the registration of partially overlapping scans. At the same time, we have demonstrated that the hybrid optimization approach in our method avoids undesired local minima, increasing the robustness of the registration process in very challenging alignment problems that involve a large number of outliers and partially overlapping surfaces. An interesting direction of research for the future is the adaptation of our hybrid optimization strategy and the underlying data structures to GPUs and similar massively parallel architectures, in order to achieve further performance improvements. Additionally, it would be interesting to explore possible extensions of our method that use salient features in order to further improve the reliability of the alignment.