描述了一种应用于视频序列的3维稠密重建方法,主要针对基于多深度图恢复3维物体时存在的多对一映射重复投影问题,提出基于误差云概念的3维定点优化方法。该方法首先通过深度矫正,增加视频序列深度图的一致性;其次利用投影过滤分类算法,把所有投影点按照多对一映射的关系,以3维空间中所有不同点为类型,进行投影点依次映射,将每一点划分在各自的误差云类中以求得投影点与3维点的对应关系;随后采用空间高斯分布求出每个误差云所恢复出来的点坐标。最后通过多边形技术对恢复的点云数据进行网格化,使其重建出精确的目标物体或场景的3维轮廓。从实验结果可以观察到,本文3维重建方法可以有效减少深度图融合多对一映射重复投影问题所带来的负面影响,使重建结果更加接近于真实物体,达到较好的效果。

This article describes a dense 3D reconstruction method based on video sequences. It mainly aims at the repeated projection problem, which exists in the 3D objects reconstruction from multiple depth maps. With the concept of an error cloud, a method called "optimization of fixed 3D points" is proposed to solve this problem. First, it uses depth rectification to improve the consistence. Then, it adopts a projection classification algorithm to classify all the projection points into the different error clouds which are characterized by the different 3D points through the many-to-one mapping relationship. At last, it computes the point coordinates for every cloud in 3D space by Gaussian distribution. Following the above, the point data are polygonized to recover the mesh of 3D objects. The final experimental results show that this method efficiently reduces the bad effect from the many-to-one mapping problem and it makes the reconstruction closer to the real object, which achieve a better effect.