| 摘要 |
<P>PROBLEM TO BE SOLVED: To determining the pose of objects for robotic picking applications. <P>SOLUTION: During pre-processing, a 3D model of the object is rendered for various poses by arranging virtual point light sources around the lens of a virtual camera. The shadows are used to obtain oriented depth edges of the object illuminated from multiple directions. The oriented depth edges are stored in a database. A camera acquires images of the scene by casting shadows onto the scene from different directions. The scene can include one or more objects arranged in arbitrary poses with respect to each other. The poses of the objects are determined by comparing the oriented depth edges obtained from the acquired images to the oriented depth edges stored in the database. The comparing evaluates, at each pixel, a cost function based on chamfer matching, which can be speed up using downhill simplex optimization. <P>COPYRIGHT: (C)2011,JPO&INPIT |
