Space Recognition as Visual Information Processing by Means of Hypothetical Reasoning

(1)视觉信息处理的假设推理（1993）我们提出了从单目图像中恢复物体三维形状的假设推理。详细地说，我们在EWS上实现了松弛优化方案，方法是在物体的三维形状上选择假设，并通过在假设下恢复的形状来验证它们。(2)通过假设和验证获得物体的描述（1993）作为(1)中提出的方法的扩展，我们研究了在当前视图中不可见的物体表面的推断。该推理采用了对物体形状的拓扑结构以及物体几何三维形状的假设。我们还研究了基于物体之间的相对位置和地板位置的假设来恢复物体在三维空间中的排列。(3)获取物体形状空间信息的传感策略（1994）我们开发了一种用于附加传感的规划视点方法，以验证由(2)中提出的方法推断出的物体不可见表面。为了使机器人到达规划的视点，我们采用视觉伺服技术对机器人的当前构型进行校正。(4)理解运动物体的动作（1993-1994）为了理解人类的协调动作，我们提出了一种利用动作的协作性和连续性知识作为一种假设，从一系列图像中识别动作模式的方法。我们还开发了一种在动作理解过程中从图像序列中提取具有显著特征的图像帧以简洁地描述动作的方法。

(1) Formulation of Hypothetical Reasoning for Visula Information Processing (1993)We have proposed hypothetical reasoning to recover the three dimensional (3D) shape of an object from its monocular image. In more detail, we have implemented the scheme of relaxational optimization on an EWS by means of the process of choosing hypotheses on the object's 3D shape and verifying them by the shape recovered under the hypotheses.(2) Acquiring the Description of an Object by Hypothesization and Verificaion (1993)As extension of the method proposed in (1), we have worked on inferring an object's surface that is invisible in the current view. The inference employs hypotheses on the topological structure of an object's shape together with those on the object's geometric 3D shape. We have also studied recovering the arrangement of objects in the 3D space based on hypotheses on the relative positions among the objects and the position of the floor.(3) Sensing Strategy to Acquire the Spatial Information of an Object's Shape (1994)We have developed a method of planning viewpoints for additional sensing to verify the objects' invisible surface inferred by the method proposed in (2). In order for a robot to reach the planned viewpoint, we employed visual servo to correct the robot's current configuration on the sensory data.(4) Understanding the Actions of Moving Objects (1993-1994)Toward understanding humans' coordinated actions, we have presented a method to recognize the action patterns from a sequence of images by employing the knowledge of cooperativeness and continuity in actions as a kind of hypotheses. We have also developed a method to extract image frames with salient features to describe the action concisely from an image sequence in the process of the action understanding.