A New Framework for 3D Point Cloud Reconstruction of Geometric Object from Multi-View Images

Abstract

 This paper proposed a new framework to enhance the 3D point cloud reconstruction from multi-view images. The accurate 3D point cloud is crucial to produce a 3D model SURF ace which can be used in many applications. There were still challenges in the existing approach to reconstruct the accurate 3D point clouds as it involves various camera orientations. This paper proposed a simplified experimental setup based on pure camera translation to capture the images from four viewpoints. The geometric object measurement is used as the ground truth data to measure the accuracy. Besides, the optical flow feature match method, perspective projection and image transformation were used to reconstruct the 3D point cloud. The result shows that all objects have an average RMSD value of 1.24mm in width and 1.26mm in depth. These values are lower than the previous method based on the images captured using known rotation, indicating higher accuracy. In future, the generated 3D point cloud can be used to create the 3D model SURF ace.