Fast restoration of geometric details of automobile castings scanned by RGB-D sensor

Abstract

The depth data of automobile castings obtained by RGB-D sensor are usually combined with noise, the classical regularization method can eliminate the noise efficiently. Yet the regularization step is too time-consuming to reconstruct the geometric details of automobile castings efficiently. Given this, we present a fast method called fast restoration of automobile castings (FRAC) to restore the geometric details of automobile castings in fast manner. First, the implicit surface data is extracted from globally aligned RGB-D images, the voxel data structure is extended to index and process the implicit surface in real time. Then, an inverse shading formula is constructed to compute TSDF (truncated signed distance field) values of casting surfaces quickly, and an objective function is designed to optimize the geometric details of casting surfaces in real time. Finally, a GPU-based Gauss–Newton solver is used to accelerate restoration of castings further. The defective casting models scanned by RGB-D sensor are quickly refined to a complete model with better accuracy. Experimental results show that with respect to the sampled automobile castings which include 359,470 points in average, the average optimization time reaches 0.66 s per frame, the average restoration time is about 6.48 s. Computing TSDF requires only about 34.8 MB GPU caches in average. FRAC is able to restore the geometric details of automobile castings in real time.