CAJ | 학술논문

基于图像的立体重建技术直接通过多幅二维图像获取物体的三维数据模型，建模自动化程度高，且不需要任何先验信息和特殊硬件支持。但对于具有精致雕刻的中国古式建筑以及非平行拍摄的大型室外场景，现有的基于图像的三维重建技术重建模型往往存在细节信息丢失、数据散乱现象，使得重建结果不够精确。针对这一问题，综合考虑模型的光照信息、纹理阴影、凹凸感等多种特征，通过给出特征候选点匹配策略及对初始点云的可靠性排序，提出了一种多特征三维稠密重建算法MFPMVS（patch with multiple features based multi-view stereopsis）。实验表明，MFPMVS算法与经典的PMVS（patch based multi-view stereopsis）算法相比，重建得到的三维点云更加密集；凹凸感较强的模型重建细节更为细腻；仰拍得到的模型重建结果中漏洞明显减少，边缘细节信息更加完整。算法能够更稳定、鲁棒地重建出物体的三维模型，具有很高的实用价值。
기우도상적입체중건기술직접통과다폭이유도상획취물체적삼유수거모형，건모자동화정도고，차불수요임하선험신식화특수경건지지。단대우구유정치조각적중국고식건축이급비평행박섭적대형실외장경，현유적기우도상적삼유중건기술중건모형왕왕존재세절신식주실、수거산란현상，사득중건결과불구정학。침대저일문제，종합고필모형적광조신식、문리음영、요철감등다충특정，통과급출특정후선점필배책략급대초시점운적가고성배서，제출료일충다특정삼유주밀중건산법MFPMVS（patch with multiple features based multi-view stereopsis）。실험표명，MFPMVS산법여경전적PMVS（patch based multi-view stereopsis）산법상비，중건득도적삼유점운경가밀집；요철감교강적모형중건세절경위세니；앙박득도적모형중건결과중루동명현감소，변연세절신식경가완정。산법능구경은정、로봉지중건출물체적삼유모형，구유흔고적실용개치。
3D reconstruction techniques based on images directly obtain 3D model information from multiple images. This kind of methods are with higher automaticity, moreover, they do not need any prior information and special hardware. But for Chinese ancient architectures with exquisite carving or the large outdoor scenes with non-parallel shooting, reconstruction results of existing 3D reconstruction techniques based on images may be not always promising because the details about modeling object are often missed or diffused. This paper considers comprehensively the multiple features of models such as lighting, texture, shadows and concavity, and proposes a novel algorithm for 3D reconstruction named MFPMVS (patch with multiple features based multi-view stereopsis) based on candidate fea-ture points mapping strategy and reliability sorting on initial cloud points. The experimental results show that, com-pared with the classical PMVS (patch based multi-view stereopsis) algorithm, the proposed MFPMVS algorithm can obtain more 3D point cloud, and the details of strong concavity model are more delicate. Meanwhile, the loop-holes of the reconstruction model with upward-shooting can be significantly reduced, and the edge information is <br> more complete. More importantly, the proposed algorithm can rebuild the 3D model of the object more stably and robustly, which means the high practicability.