SIFT and ICP in Multi-view based Point Clouds Registration for Indoor and Outdoor Scene Reconstruction

Abstract

3D reconstruction requires efforts in solving the fundamental problems such as accuracy, wholeness, and acquisition method. To achieve a rigid and whole shape, 3D object reconstructed from many different views should be registered. The registration process aims to combine and align pieces of the same object which are having different orientations. The registration process requires some steps started from data acquisition, feature extraction and matching, essential matrix calculation and decomposition, coarse registration, and refinement. We propose the use of SIFT (Scale Invariant Feature Transform) algorithm in feature extraction and matching process to find point correspondences on each indoor and outdoor multi-view scene images, along with other followed registration steps. The search of point correspondences is crucial, since it appears as basis in essential matrix calculation that lets us obtain the orientation value between all reconstructed pieces. The complete registration process ends up with refinement using ICP (Iterative closest point) towards the coarse registration results. It is known at the end of research that SIFT works really good in feature extraction and matching for supporting the point clouds registration. According to SIFT and SURF (Speeded Up Robust Feature) comparison table, SIFT could extract relatively 2.09 times more features than SURF with relative calculation time of 152.19 times longer than SURF. In refinement process using ICP, the average error is decreased by 64.97% in white car dataset, 8.88% in kitchen 1 dataset, and 87.12% in kitchen 2 dataset.