| 摘要 |
<P>PROBLEM TO BE SOLVED: To execute efficient bootstrap of artificial vision recognition even in an unknown environment where an object without being previously grasped appears. <P>SOLUTION: This system is so structured that a couple of synchronized time steps are continuously set; two continuous images in each couple of the synchronized time steps separated by a delay time τ<SB>0</SB>are compared with each other; distances between the respective pixels of the two continuous images are calculated by considering characterization of motion of the object to obtain a delta image Δ<SB>t</SB>. The system is so structured that a characteristic part is extracted from the delta image Δ<SB>t</SB>to find a dynamic patch candidate P<SB>t</SB>to be compared with a dynamic patch previously stored in a repertory; the dynamic patch D<SB>i</SB>having the most similarity is selected from the repertory, or the candidate P<SB>t</SB>is added in the repertory when no dynamic patch having high similarity exists and thereby the patch D<SB>i</SB>is obtained and stored; and the dynamic patches D<SB>i</SB>stored in the repertory are integrated in terms of time to detect and store a stable set of active dynamic patches showing characteristics of observed and reoccurring motion or events. <P>COPYRIGHT: (C)2004,JPO |
