Experimental Partnership - Real-Time Computer Vision Based Spatial Mapping and Referencing for Minimally Invasive Surgery

badrinath RoysamRensselaer理工学院实验合作伙伴关系-基于实时计算机视觉的微创手术空间映射和参考目标是为新一代高性能微创手术器械开发核心计算机视觉技术。具体应用在眼科。开发这项技术需要面对和克服困扰计算机视觉系统的几个基本障碍。这些问题包括手术过程中图像数据质量差，生物组织建模困难，眼睛的大规模和不可预测的运动，需要长时间的极端精度，需要实时实施中可预测的响应，以及需要整个系统的透明度。计算机视觉技术将通过“空间感知”来区分这些乐器。这意味着它们包括空间映射和空间引用的功能。空间映射是建立和维护一个无缝的，广域地图（马赛克）的三维区域的手术利益及其周围的问题。该地图将根据诊断探索期间获得的图像构建，并构成绝对的、因此可验证的空间参考的基础。测绘工作的新方面包括：(1)一种快速递归技术，用于跟踪细长结构并检测其分支和交叉点；(2)一种基于特征的图像到图像匹配算法，该算法使用高阶变换模型和层次鲁棒估计技术；(3)同时对所有图像到马赛克转换进行联合优化；(4)尽管固有的无法校准成像系统，但仍可以对视网膜表面进行三维重建。这些技术在眼科以外的其他领域都很有用。

EIA-0000417Badrinath RoysamRensselaer Polytechnic InstituteExperimental Partnership-Real Time Computer Vision Based Spatial Mapping and Referencing for Minimally Invasive SurgeryThe goal is to develop core computer vision technology for a new generation of highly capable instruments for minimally invasive surgery. The specific applications are in opthamology. Developing the technology requires facing and overcoming several fundamental barriers that have plague computer vision systems. These include poor quality image data during surgery, difficult to model biological tissues, large scale and unpredictable motions of the eye, a need for extreme accuracy over long duration, a need for predictable response in a real-time implementation, and a need for transparency of the system as a whole.The computer vision technology will distinguish these instruments by making them "spatially-aware". This means they include capabilities for spatially mapping and spatial referencing. Spatial mapping is the problem of building and maintaining a seamless, wide-area map (mosaic) of the three-dimensional region of surgical interest and its surrounding. This map will be constructed from images acquired during diagnostic exploration and forms the basis for absolute and hence verifiable spatial referencing.Novel aspects of the work in mapping include: (1) a fast, recursive technique for tracing elongated structures and then detecting their branching and cross-over points, (2) a feature-based image-to-image matching algorithm that used a high-order transformation model and a hierarchy of robust estimation techniques, (3) a joint optimization of all image-to-mosaic transformations simultaneously, and (4) a 3-D reconstruction of the retinal surface despite an inherent inability to calibrate the imaging system. These techniques are each useful in other domains besides ophthalmology.