TomoSAR II: 3-D Semantic Scene Interpretation of Tomographic SAR Data

这个项目是我们之前项目“层析合成孔径雷达（TomoSAR）数据的信息提取方法”的后续项目。我们的目标是专注于算法的发展，以获得包含在层析SAR数据中的三维信息的自动语义解释。为了实现这一目标，我们将首先结合之前建立的二维图像分类结果，对利用丰富的三维描述符的层析图定义一种三维监督分类方法。该分类器将适应于识别“连续”的物体，如森林、草地、街道和田野。为了识别“离散”物体，如建筑物和其他人造结构，我们将提出两种互补类型的探测器。第一个被称为“知觉分组”，将考虑几何原语的分组，这得益于格式塔心理视觉理论的启发。第二种方法是通过描述关键点特征组的高级模型来定位和估计物体的三维姿态。我们的新检测方法将依赖于生成一个目标数据库，这得益于一种适应的TomoSAR模拟方法的发展，解决了SAR训练数据的缺乏。由于基于成本的公式，来自这些分类器和检测器的所有信息将被组合到一个单一的场景解释框架中。这种表示将在反馈回路中被利用，以改进层析图和单个对象提取的质量。循环将通过更新来自前面步骤的输出的语义解释来结束。此外，我们将继续研究空间自适应协方差估计，以解决单视图数据过滤问题，并解决过滤数据的模型阶估计问题。这些步骤将集成到TomoSAR处理链中，从而导致我们最终的场景解释。

This project is a follow-up to our previous project entitled “Information extraction methods for Tomographic Synthetic Aperture Radar (TomoSAR) data”. Our objective is to focus on the development of algorithms to obtain an automatic semantic interpretation of the 3-D information contained in tomographic SAR data.To achieve this goal, we will first define a 3-D supervised classification method for tomograms exploiting rich 3-D descriptors, combined with the result of previously established 2-D image classification. This classifier will be adapted to recognize “continuous” objects such as forest, grassland, street and field. To identify “discrete” objects such as buildings and other man-made structures, we will propose two complementary types of detectors. The first one called “perceptual grouping” will consider the grouping of geometric primitives thanks to features inspired from the Gestalt psycho-visual theory. The second one is aimed at localizing and estimating the 3D pose of objects thanks to high-level models describing groups of keypoint features. Our new detection approaches will rely on the generation of an object database obtained thanks to the development of an adapted TomoSAR simulation method, addressing the lack of training data for SAR. All the information from these classifier and detectors will be combined into a single scene interpretation framework thanks to a cost-based formulation. This representation will be exploited in a feedback loop to refine the tomograms and the quality of individual object extraction. The loop will be closed by updating the semantic interpretation from the outputs of the previous steps.Additionally, we will continue to work on spatially adaptive covariance estimation to solve the problems of single-look data filtering and address the issue of model order estimation for filtered data. These steps will be integrated into a TomoSAR processing chain leading to our final scene interpretation.