基于稳健估计的球形全景影像稠密深度图生成方法研究

Panoramic dense depth map is very important for the integrity of the full-field three-dimensional scene reconstruction. There are mainly two kinds of difficulties for restoring panoramic dense depth map: The one is the effects of lighting changes, perspective distortion, weak textured area, occlusion, tilt surface, etc. The other is the effects of panoramic projection geometry, rotation transformation and scaling relations. The existing research of panoramic depth map has not yet fully considered these two types of factors, mainly some parts of the panorama dense depth map generation. So, this project will do research on constructing robust dense depth map generation to reduce the interference of various factors. There are mainly three problems to be solved as follows: (1) Research and construct a kind of matching method based on global multi - clue fusion and a gross error elimination method based on multi-constraints, completing panoramic matching and orientation parameters optimization. (2) Analyze and construct a kind of panoramic area segmentation and semantic recognition method based on adaptive multilevel panoramic epipolar line, and complete sparse feature point matching diffusion. (3) Research and construct a kind of multi-restrictions energy optimization and plane estimating method based on higher-order Markov Random Field to complete the panoramic dense depth map, and construct high fidelity texture mapping based on energy function optimization. This research can provide the integrity of 3D data, which will have great value to the development of virtual augmented reality and 3D GIS.

全景影像稠密深度图对三维场景重建的完整性具有重要意义。生成全景稠密深度图主要面临两方面的影响：一是光照变化、透视畸变、弱纹理区域、遮挡、倾斜表面等传统因素；二是全景影像几何投影、旋转变换及缩放关系等特定因素。现有的全景深度图研究尚未充分考虑这两方面影响，大多是全景深度图生成的部分环节。鉴于此，本课题探索构建基于稳健估计的球形全景影像稠密深度图生成方法，以减少各种因素干扰。项目主要解决以下三个问题: (1)研究构建全局多线索融合匹配方法与多约束条件的粗差剔除方法，完成立体全景特征匹配与定向参数优化；(2)研究构建语义区域识别的全景核线自适应多级松弛扩散方法，完成稀疏特征点匹配扩散；(3)研究构建基于多约束高阶马尔科夫随机场能量框架的全景平面统计推断方法，实现全景稠密深度图生成，并基于能量优化构建高保真全景纹理映射。本项研究可提供完整性的三维数据源，对推动虚拟现实、三维GIS发展具有重要价值。

全景影像稠密深度图对三维场景重建的完整性具有重要意义。生成全景稠密深度图主要面临两方面影响：一是光照变化、透视畸变、弱纹理区域、遮挡、倾斜表面等传统因素；二是全景影像几何投影、旋转变换及缩放关系等特定因素。现有的全景深度图研究尚未充分考虑这两类因素影响，仅是全景稠密深度图生成中的部分环节。主要研究内容：(1) 建立全景特征匹配及其立体影像定向参数优化方法；（2）建立全景核线自适应多级松弛扩散方法；（3）构建全景平面统计推断方法。课题以全景影像稠密深度图构建为研究对象开展相关研究。本项目发表文章14篇，其中SCI收录期刊7篇，EI收录期刊3篇，中文核心3篇，国际会议1篇，获得云南省科学技术自然科学奖三等奖1项，项目成员2人获得云南省中青年学术和技术带头人后备人才，1人获得中青年学术和技术带头人后备人才，2人晋升为三级教授，1人晋升为教授，2人晋升为副教授，1人获得博士学位。本项研究可提供完整性的三维数据源，对推动虚拟现实、三维GIS发展具有重要价值。