摘要
由于匹配信息弱或噪声影响,深度计算精度难以保证,故深度图融合是多目立体视觉3维重建中的关键部分。为此,提出一种基于置信度的抗噪融合算法。该方法首先对每幅深度图进行修正,利用一致性检测剔除大多数错误点并填补某些空洞。其次,通过保留那些在自身邻域内具有最高置信度的3维点以删除冗余。最后,将深度图反投影到3维空间,采用迭代最小二乘法进一步优化3维点并剔除离群点。通过在标准测试数据集上与其他算法比较,验证了该方法的有效性。
Due to the weakness of match information and influence of noise, the calculation precision of depth cannot be guaranteed. Therefore the fusion of multiple depth maps is a typical technique for multi-view stereo （ MVS ） reconstruction. An antinoise fusion method that took advantage of the confidence of 3D points was introduced. This method performed a refinement process on every depth map to enforce consistency over its neighbors, which could remove most errors and fill many holes simultaneously. After： refinement, it deleted redundancies of every point by retaining the point that its confidence was maximal in its neighbors. Finally, it obtained a point cloud by merging all depth maps and used an iterative least square algorithm to further eliminate the noise points. The quality perform- ance of the proposed method was evaluated on several data sets and compared with other algorithm.
出处
《四川大学学报（工程科学版）》
EI
CAS
CSCD
北大核心
2016年第4期101-106,共6页
Journal of Sichuan University (Engineering Science Edition)
基金
国家自然科学基金资助项目(61571313)
四川省科技厅资助项目(2014HH0048)
关键词
多目立体视觉
3维重建
深度图融合
置信度
迭代最小二乘法
multiple view stereo
3D reconstruction
fusion of depth maps
confidence
iterative least square algorithm
