多视照片全局约束下的Kinect在线室内三维测图技术研究

Kinect as an outstanding representative of new RIM (Range Imaging) camera has different data features. Taking Kinect as a starting point, the development of RIM camera 3D mapping technology is significant for indoor navigation application. Aiming at the existing problems of Kinect accumulated error with continuous location, this paper put forwards a new idea that using multi-view photos as global constraints with online Kinect mapping. In advance, we take multi-view photos with wide field information to build a closed-loop model of global constraint network. In support of this network, we make breakthroughs on coarse positioning and real-time joint adjustment for fine position. The objective is to realize the precise and controllable online 3D mapping. Through the research of construction method of global constraint network, quality inspection based Kinect failure positioning judgment, heuristic Kinect relocation of auxiliary feature point support and real time joint adjustment with Kinect and multi-view photos, this paper will form the technology chain of "positioning failure judgment-assisted coarse positioning-adjustment precision positioning", which can not only enrich the RIM camera measurement system theory and method, but also has important application value in rapid 3D mapping for complex indoor environment.

Kinect作为新型RIM（Range Imaging）相机的杰出代表，具有独特的数据特点，以其为切入点，发展RIM相机在线三维测图技术，对室内导航具有重要意义。现有Kinect技术进行大场景三维测图时，相机连续定位易出现累计误差，导致模型重建不精确。针对该问题，本课题以"多视照片全局约束，Kinect局部精细建模"为总体指导思想，提出利用高清多视照片的宽视场信息预先构建闭环模式的全局约束网，来对Kinect建模过程进行全局控制，通过突破Kinect局部精细模型在全局约束网中的粗定位和实时平差关键技术，实现精确可控的在线测图。通过研究全局约束网构建方法、基于质量检查的Kinect定位失效判断、辅助定位点支持的启发式重定位、Kinect与高清照片的实时平差算法，形成"定位失效判断-辅助粗定位-平差精定位"技术链条，丰富RIM相机测图理论与方法体系，提升室内复杂环境下快速三维测图的应用价值。

目前，室内场景三维测图常采用昂贵的LiDAR激光扫描仪，需提前设置拍摄站点和标靶，扫描前期准备工作较长。本项目针对现有技术中离线测图存在的不足，本项目提供了一种基于Kinect和流媒体技术的实时三维测图方法，具有轻便、灵活、成本低廉；实时测图；前期工作耗时少，扫描效果好等优点。课题组按照计划对室内三维地图构建进行了深入研究，研究了基于FNEA多尺度聚簇的初始约束网构建，基于层次图模型的实时联合平差算法，带空间范围约束的特征点最邻近查询索引等关键技术，构建了一套有利于的Kinect实时三维测图设备，初步形成“多视照片全局约束，Kinect局部精细建模”技术体系，可实现面向LiDAR、照片和Kinect数据的室内三维场景的拼接合并。项目发表SCI论文1篇，获得国家发明专利授权3项，EI论文1篇，培养博士生1名，硕士生1名，组织召开国际会议1次，参与国际会议2次，为全息位置地图三维构建提供了框架支撑，并将相关技术扩展到了室内定位和室内三维行为识别方面，在武汉市公安局犯罪现场重建进行了初步试用。项目后续研究成果将持续发表，目前还有1篇SCI论文和1篇EI论文进入编辑部终审，待出版专著1部。