自己評価を伴うアクティブビジョン

Abstract

Active visionfor constructing a three-dimensional model of the environment from images requires a robot to control its own motion. Since noise exists in images, the information they provide is not always complete. If the motion is too small, images can provide only2-D informationwithout any depth clues. As the motion increases, we obtain incomplete 3-D information, which we call2.5-D information. After the motion becomes sufficiently large, we obtain complete3-D information. We give a geometric interpretation to these transitions by viewing the problem asmodel fittingof a manifold in an abstract data space. We also derive a decision rule based on thegeometric AIC. This rule can be used as a means ofself-evaluationfor testing if the robot motion is sufficient for structure-form-motion analysiswithout involvinganyempirically adjustable thresholds. To demonstrate this, we give examples using synthetic and real-image data.