PHASER: A Robust and Correspondence-Free Global Pointcloud Registration

Abstract

We propose PHASER, a correspondence-free global registration of sensor-centric pointclouds that is robust to noise, sparsity, and partial overlaps. Our method can seamlessly handle multimodal information, and does not rely on keypoint nor descriptor preprocessing modules. By exploiting properties of Fourier analysis, PHASER operates directly on the sensor's signal, fusing the spectra of multiple channels and computing the 6-DoF transformation based on correlation. Our registration pipeline starts by finding the most likely rotation r ∈ SO(3) followed by computing the most likely translation t ∈ R 3. Both estimates, r, and t are distributed according to a probability distribution that takes the underlying manifold into account, i.e., a Bingham and a Gaussian distribution, respectively. This further allows our approach to consider the periodic-nature of r and naturally represents its uncertainty. We extensively compare PHASER against several well-known registration algorithms on both simulated datasets, and real-world data acquired using different sensor configurations. Our results show that PHASER can globally align pointclouds in less than 100 ms with an average accuracy of 2 cm and 0.5 • , is resilient against noise, and can handle partial overlap.

DPCN++: Differentiable Phase Correlation Network for Versatile Pose Registration

“…The globally convergent correspondence-free method mainly employs the idea of phase correlation. B ülow et al [5] and PHASER [3] both utilize spherical and spatial Fourier transforms to estimate the relative pose using correlation [57] in the spectrum. The global convergence lies in the correlation, which is an intrinsically exhaustive search, but can be evaluated efficiently via decoupling in the spectrum.…”

“…Inspired by these methods, we introduce a differentiable version of phase correlation to enable end-to-end learning based on a globally convergent solver. In contrast to [5], [3], our framework achieves better registration performance and is applicable to versatile pose registration tasks by learning from data. Recently, with the progress of geometric deep learning, Zhu et al [77] apply SO(3)equivariance embedding for feature learning.…”

ARCS: Accurate Rotation and Correspondence Search

“…Modern sensors have brought the classic Wahba problem [86], or slightly differently the Procrustes analysis problem [36], into greater generality that has increasing importance to computer vision [39, 57], computer graphics [65], and robotics [14]. We formalize this generalization as follows.…”

“…For Problem 1 or its variants, there is a vast literature of algorithms that are based on i) local optimization via iterative closest points (ICP) [15,23,76] or graduated nonconvexity (GNC) [2,87,96] or others [26,46,68], ii) global optimization by branch & bound [21,24,57,61,62,73,80,92], iii) outlier removal techniques [19,71,72,79,91], iv) semidefinite programming [44,65,81,88,90], v) RANSAC [33,58,59,82], vi) deep learning [5,10,25,42], and vii) spherical Fourier transform [14]. But all these methods, if able to accurately solve Problem 1 with the number k * of inliers extremely small, take Ω(mn) time.…”

SemSegMap- 3D Segment-Based Semantic Localization

“…Many PC registration methods, out of which Iterative Closest Point (ICP) [18] is the most well-known one, require a good pose prior and are not suited for global localization. While global registration methods exist that work beyond the local context [19], [20], they still require storing at least parts of the PC data. This can be partially mitigated by only extracting compact descriptors during map building and localization.…”