An automatic and robust point cloud registration framework based on view-invariant local feature descriptors and transformation consistency verification

This paper presents an automatic and robust framework for simultaneously registering pairwise point clouds and identifying the correctness of registration results. Given two partially overlapping point clouds with arbitrary initial positions, a view-invariant local feature descriptor is utilized to build sparse correspondence. A geometry constraint sample consensus (GC-SAC) algorithm is proposed to prune correspondence outliers and obtain an optimal 3D transformation hypothesis. Furthermore, by measuring the similarity between the estimated local and global transformations, a transformation consistency verification method is presented to efficiently detect potential registration failures. Our method provides reliable registration correctness verification even when two point clouds are only roughly registered. Experimental results demonstrate that our framework exhibits high levels of effectiveness and robustness for automatic registration.