A Study of Effective Recognition of Environment by Integrating Global and Local Visions

点击翻译按钮获取中文摘要

Last year, we exploited a new omni-directional image sensor in which a hyperboloidal mirror is aligned axis-to-axis over the input-lens of camera. Global models or maps of the environment surrounding a robot were aquired by loading with this sensor, and the planning and the revision of a moving path were succesfully achieved. Our new achievement in this year is as follows.[1] Detection of a Target Object :A method to detect a predefined target object is proposed. For example, the target can be (1) the object to work on, (2) landmarks on the path to a given goal, (3) obstacles which disturb the robot from moving, (4) some suspicious object, or (5) a moving one against which the robot may collide, and so on. In our method, specialized modules are installed, each assigned the task of finding objects in one of these categories. Each module looks around the overall surrounding area, and if detecting any candidate of the target, run the local-vision module to look into its fine feature.[2] Recognition of Fine 3D-structure of a Detected ObjectThe sensor, focusing attention to an object detected by the global vision, recognizes the fine 3D-structure of the object. Since the abstract structure is already known by the global sight, another camera which realizes a local vision is shifted so that the target comes into the center of the image, and then zoomed until the size of the object's image grows as large as the vision-frame. The focus of the lens is also arranged to make the image clear. Based on the input from plural directions, the accurate 3D-structures of target objects were obtained eventually.[3] Experimental Evaluation of the Integral Vision SensorWe developed an image sensor in which the global and the local visions are integrated. Employing this new sensor, objects in real environments suchas the corridor of our university or our laboratory were recognized succesfully.