Recognizing and Locating Partially Occluded Objects by Boundary

• Main Authors: Bor-Ren Liaw, 廖博仁
• Other Authors: Du-Ming Tsai
• Format: Others
• Language: zh-TW
• Online Access: http://ndltd.ncl.edu.tw/handle/07329961441889723796

碩士 === 元智大學 === 工業工程研究所 === 82 === The problem of recognizing partially occluded parts is of considerable interest and challenge in the field of industrial automation. While it is possible to employ carriers or pallets to separate or prearrange the parts for easy recognition, a vision system which can recognize the parts even though they may be partially overlapped and in random positions in much flexible. An occluded object can be due to the occurrence of touching and overlapping objects, or the occurrence of defects. The polygonal approximation of contours is the most frequently used technique to represent the shape of objects, and the line segments of the observed shapes are matched with the segments of model objects. However, this method is sensitive to noise and gives satisfactory results only for polygonal objects. In this research, we consider the problem of identifying and locating partially occluded objects lying on a flat surface, and the contours of objects can be arbitrary piecewise smooth curves. A novel shape representation is developed to describe the boundaries of objects. Given a boundary point on the curve, we use the least-square circle fitting method to estimate the center of circle for a small curve segment in the neighbor of Pi. A vector Ni connecting the estimated center and point Pi is formed to represent the geometric features of the boundary point. The magnitude of the vector (or, the radius of the estimated circle) represents the curvature of the curve at point Pi, and the vector direction represents the normal to the curve at point Pi. Then, a two- stage Hough transform is used for the recognition of occluded objects. The vector N is used as the shape signature that gives geometric constraint information in the voting stage of the Hough transform. It eliminates noise and false matching boundary points. The matching correctness and reliability and the accuracy of the object pose (rotation angle and translation) are improved accordingly.