| Abstract: | Abstract: The segmentation and localization of overlapped apples in nature scenes are the key factors for the efficiency of picking robot's. To achieve successful recognition of occluded apples on the basis of K-means color clustering algorithm, a convex hull based concave point detection algorithm is presented. K-means color clustering algorithm is used to detect the apples, in which the parameter K is 3 (that is, the image is clustered to 3 different kinds such as the apples, the leaves and branches). The depressed regions are obtained by subtracting the binary image from its convex hull image after applying the color-clustering algorithm. By using image opening algorithm, the regions that areas are less than half the biggest region are removed. The remained depressed regions are used to detect the point of intersection of overlapped apples, and then the concave point detection is transformed to salient point detection, which reduces the complexity of classical concave point detection algorithm. Some split criteria are given as: 1) The length of split line should be short; 2) The direction of the detected points should be opposite, which means that the detected points from the same region should not be used to split an object; 3) The split contours must be a major arc. By using these given criteria, the detected concave points are matched to realize the segmentation of overlapped apples. The overlapped apple regions are unable to use, and these parts are incompletely segmented. For these defectively segmented apples, the spline interpolation technique is used to reconstruct the apples: 1) The centroid of segmented region is used as the starting point and the points on the contour are transformed from orthogonal coordinate to polar coordinate; 2) Rearranging the point by the value of polar angles, from small to big; 3) The rearranged pole diameters are interpolated by using spline interpolation technique; 4) The original polar angels and interpolated polar diameters are retransformed from polar coordinate to orthogonal coordinate to get smoothed, complete contour of overlapped regions. To validate the effectiveness of the algorithm, 3 different kinds of simulating examples (2 overlapped objects, 3 overlapped objects and 5 overlapped objects) and 20 overlapped apples in nature scenes are tested, the average localization error is 14.15%, and the average overlap ratio is 85.08%. The experimental results show that the convex hull based occluded objects segmentation algorithm has preferable performance, and it is feasible and valid for overlapped apple segmentation in nature scenes. |
