重叠苹果果实的分离识别方法

Separation and recognition method for overlapped apple fruits

针对苹果采摘机器人重叠果实识别误差较大的问题,设计了一种分离识别方法。首先在苹果图像分割获取其二值果实区域的基础上,基于横、纵投影图实现重叠形态果实的判别,而后基于边缘曲线通过SUSAN算法检测果实轮廓上的角点,再通过迭代腐蚀和瓶颈准则挑选重叠果实的分离点,并采用Bresenham算法连接分离点实现重叠果实的分离。提取分离果实边缘曲线的有效轮廓后,通过改进的随机Hough算法拟合果实圆心及半径。最后选择15幅重叠果实区域二值图像,通过不同角点检测计算结果的比较,验证了SUSAN算法相比于其他角点检测方法更为有效;通过改进的随机Hough算法识别11幅图像中的21个果实,其圆心相对误差平均值、半径相对误差平均值和相对偏差平均值分别为6.90%、4.12%和6.07%,比传统Hough算法分别低4.03、2.75、1.14个百分点,说明改进的随机Hough算法得到的拟合圆更接近实际苹果果实区域。

In order to recognize the overlapped apple fruits accurately, a method of separation and recognition of overlapping fruits for apple picking robot was designed. Firstly, on the basis of obtaining the binary image of apple fruit area, the fruit of the overlapping form was distinguished by using the transverse and longitudinal projection, and then the corner points were detected by the SUSAN algorithm based on the fruit edge curve, and then the separation points of the overlapped fruits were selected by the iterative corrosion and bottleneck criteria, and Bresenham algorithm was used to connect separation points to achieve the separation of overlapping fruit. After extracting the effective contour of separating fruit edge curve, the center and radius of fruit were fitted by the improved random Hough algorithm. Finally, 15 binary images of overlapping fruit areas were selected to verify that SUSAN algorithm was more effective than other corner detection methods by comparing the results of different corner detection. The improved random Hough algorithm was used to identify 21 fruits in 11 images. The average values of relative error of the center, the relative deviation of the radius and relative deviations were 6.90%, 4.12% and 6.07% respectively, which were 4.03%, 2.75% and 1.14% lower than those acquired by the traditional Hough algorithm. This indicated that the fitting circle obtained by the improved random Hough algorithm was closer to the actual apple fruit area.