Two-Stage Multi-Camera Constrain Mapping Pipeline for Large-Scale 3D Reconstruction

Over the last decade, Augmented reality (AR) has witnessed a significant increase in popularity largely. Consequently, as a central component of these applications, large-scale 3D reconstruction has gained more and more attention. One of the promising solutions is using the camera and the Structure from Motion(SfM) method, especially with the advent of multi-camera devices. However, most of these methods are susceptible to failure in some high-traffic commercial street scenarios due to lots of moving objects and texture-free scenes. Besides, these methods treat each image as a separate source of information, which significantly degrades the performance of the multi-camera acquisition systems now commonly used. Hence, this paper presents a two-stage reconstruction pipeline to obtain an accurate large-scale 3D map in the central business district with multi-camera devices. In the first stage, we enhance the robustness of the incremental SfM method by using the pose prior obtained from the multi-camera visual-inertial system(VINS) and advance the accuracy of the reconstruction result in the second stage by performing the global optimization considering the multi-camera extrinsic constraints. Experiments show that our method yields more complete and accurate maps than its counterpart.