Description of Three-Dimensional Environment by Reliable Stereo Vision

(1)[可靠立体]为了解决在两幅立体图像中寻找对应点的中心问题，我们提取了两幅图像中的特征(边缘)，并根据边缘的相似性进行匹配。将图像划分为小区域(窗口)，并在每个窗口中获得匹配的候选图像。从配对的视差直方图来评价匹配的可靠性。提出了一种多阶段匹配方法，首先对可靠的边缘进行匹配，然后利用先前的结果对不可靠的边缘进行处理，然而，如果窗口包含不连续的曲面，则可能得不到对应或可能包含错误。由于不连续曲面会在图像中产生T形连接，因此需要检测不连续曲面的候选并重复匹配过程，尽管大多数传统的立体视觉假设摄像机的位置是已知的，但这在实际情况下并不成立。为了降低错位误差，提出了一种自标定方法。(2)[可靠立体场景描述]从立体视觉获取的三维信息中提取曲面的不连续、突变等特征。几何描述是由特征生成的。我们开发了一种方法，通过将距离数据组合成灰色或彩色图像来生成丰富的描述。所得到的描述的精度取决于三维信息的误差。我们提出了一种评估描述可靠性的方法，并通过实验进行了验证。使用关于对象的知识从描述中识别场景。

(1) [Reliable stereo] To solve the central problem of finding corresponding points in a pair of stereo images, we extracted features (edges) in the both image and tried to match them according to the similarity of the edge. The image is divided into a small regions (windows) and the matching candidates are obtained in each window. The reliability of matching is evaluated from the histogram of the disparity of the pairs. A multistage method is developed in which matching is tried for reliable edges first, and the less reliable edges are processed later using the previous result.However, correspondence may not be obtained or may include errors if a window contains discontinuous surfaces. Since a discontinuous surfaces produce T-junctions in an image, the candidates of discontinuous surfaces are detected and matching process is repeated.Although most conventional stereo visions assume that the position of cameras are known, this does not hold in an actual case. To decrease the error caused by miss alignment, a self calibration method is developed.(2) [Scene description by reliable stereo] From three-dimensional information obtained by the stereo vision, features are extracted such as discontinuity and abrupt change of surfaces. The geometric description is generated from the features. We developed a method of generating a rich description by combining the range data to gray or color images. The precision of the obtained description depends on the error of three dimensional information. We developed a method of evaluating the reliability of the description, and verified by experiments. A scene is recognized from the description using knowledge about objects.