Real-time 3D features reconstruction through monocular vision

Abstract

A fast and interactive implementation for cam- era pose registration and 3D point reconstruction over a physical surface is described in this paper. The method (called SRE—Smart Reverse Engineering) extracts from a continuous image streaming, provided by a single cam- era moving around a real object, a point cloud and the camera's spatial trajectory. The whole per frame procedure follows three steps: camera calibration, camera registra- tion, bundle adjustment and 3D point calculation. Cam- era calibration task was performed using a traditional approach based on 2-D structured pattern, while the Opti- cal Flow approach and the Lucas-Kanade algorithm was adopted for feature detection and tracking. Camera regis- tration problem was then solved thanks to the Essential Matrix definition. Finally a fast Bundle Adjustment was performed through the Levenberg-Marquardt algorithm to achieve the best trade-off between 3D structure and cam- era variations. Exploiting a PC and a commercial web- cam, an experimental validation was done in order to verify precision in 3D data reconstruction and speed. Practical tests helped also to tune up several optimization parame- ters used to improve efficiency of most CPU time consum- ing algorithms, like Optical Flow and Bundle Adjustment. The method showed robust results in 3D reconstruction and very good performance in real-time applications.