Outdoor World Modeling by Intelligent Integration of Multi-Visual Information

在本项目中，我们提出了一种场景解释方法，该方法利用场景中物体的知识和单个感官数据的属性，根据感官数据处理的中间结果，将多视觉感官数据（如强度和距离图像）动态地整合到场景描述中。在第一年，我们发现利用多视觉信息对一类物体进行几何建模对于该类物体的提取和场景中物体之间的空间关系推理是有用的。今年，我们扩展了我们的方法，使它可以适用于各种类型的对象类。也就是说，我们开发了一个模型驱动的空间推理系统，从背景中提取各种物体，并确定这些物体和未探索区域的几何结构。研究结果如下：(1)对于立体彩色图像，基于特征的立体匹配方法获得的距离数据是稀疏的，不足以对场景进行分割。然后，利用在视差空间中提取的平面斑块，利用对场景结构有用的颜色信息对区域分割结果进行恢复；因此，大致恢复了场景结构，但需要更准确的距离信息来确定场景结构的细节。(2)对于一对密集距离数据和一幅强度图像，我们从距离数据估计的平面斑块的不确定性中确定参数模型的参数。研究表明，不确定性取决于两种误差源：可能误差和系统误差。我们揭示了它们之间的关系，并将其应用于物体识别问题。

In this project, we proposed a method of scene interpretation which dynamically integrates multi-visual sensory data such as intensity and range images into the scene descriptions according to the intermediate results of the sensory data processings that are obtained by using the knowledge of objects in the scene and the properties of the individual sensory data. In the first year, we found that geometric modeling of one object class by using multi-visual information was useful for extraction of the objects in this class and for reasoning of spatial relationships between objects in the scene. This year, we extended our method so that it could be applicable to various kindsof object classes. That is, we developed a model-driven spatial reasoning system which extracted various kinds of objects from the background and determined the geometric structures of these objects and unexplored regions as well. Research results are as follows:(1) In the case of stereo-color images, the range data obtained from a feature based stereo matching method are sparse, therefore they are not sufficient for segmentation of the scene. Then, we utilized the result of region segmentation by using the color information which was useful to recover the scene structure in terms of planar patches extracted in the disparity space. As a result, the scene structure was roughly recovered, but more accurate range information is needed to determine the details of the scene structure.(2) In the case of a pair of dense range data and an intensity image, we determined the parameters of the parametric model from the uncertainties of planar patches estimated from the range data. It has been shown that the uncertainty depends on two kinds of error sources: probable and systematic errors. We revealed their relationships and applied it to the problem of object recognition.