苹果采摘机器人快速跟踪识别重叠果实

为解决采摘机器人在运动状态下对重叠果实的识别问题,减少采摘过程处理的时间,对重叠果实的快速跟踪识别进行了研究。首先,对采集到的第1幅图像进行分割并去噪,之后通过计算圆内的点到轮廓边缘最小距离的极大值确定圆心的位置,计算圆心到轮廓边缘距离的最小值确定半径,通过圆心与半径截取后续匹配的模板,经试验证明该算法能较准确地找到重叠果实的圆心与半径。然后,确定连续采集的10幅图像的圆心,根据每幅图像圆心的位置对机器人的运动路径进行拟合、预判、综合半径与预判路径确定下一次图像处理的范围。最后,采用快速归一化互相关匹配对重叠果实进行匹配识别。试验证明,经过改进后的算法匹配识别时间与原算法相比,在没有进行预判的情况下匹配识别的时间为0.185 s,经过预判之后,匹配时间为0.133 s,减少了28.1%,采摘机器人的实时性得到了提高,能够满足实际需求。该研究可为苹果等类球形重叠果实的动态识别提供参考。

Fast tracking and recognition of overlapping fruit for apple harvesting robot

Abstract: Fruit identification and location are the primary tasks and difficulties for fruit-harvesting robot, whose accuracy closely relates to the efficiency of robot. Many scholars, at home and abroad, have carried out a large number of researches on overlapping fruits and have achieved some initial results. However, these studies are limited to static conditions, which are not available for moving state. In recent years, some scholars have studied dynamic identification of harvesting robot and have proved that image processing time could be effectively reduced by using the correlation of a series of images, but the recognition of overlapping fruits is not included. In order to solve the problem of identification of overlapping fruits in moving state and real-time harvesting of robot, the research of fast tracking identification of overlapping fruits is conducted. Firstly, the collected image should be segmented and denoised by mophology method, then the location of the center is determined by calculating the maximum value of minimum distance between the edge of contour and points within the circle. Four scanning directions are defined in order to improve the timeliness; in direction A, the points are scanned from left to right and from top to bottom; in direction B, from right to left and from top to bottom; in direction C, from right to left and from bottom to top; in direction D, from left to right and from bottom to top. The radius is determined by calculating the distance from the center to the edge contour, but in the case of overlapping fruit, the distance from the center to the edge may be the distance to another apple, so a new method is used to avoid this situation. The distances from the center to the edge in different directions are calculated and the minimum distance is used as the final radius. After that, the template of the follow-up match is intercepted by the center and radius. The experiment proves that the algorithm can accurately find the center and radius of overlapping fruit. Then, the centers of images which are collected continuously should be determined. According to the center position of each image and the sampling time of the robot, the motion path of the robot is fitted and predicted. In this experiment, seven-order curve fits robot trajectory well. The processing scope of the next image is determined by radius and anticipated path. Finally, fast normalized cross-correlation match is used to identify overlapping fruit. Fast normalized cross-correlation method is simple and eliminating the issue of light sensitivity, so it is applied to apple images under different light intensities, besides, its matching result is also accurate when the image is slightly displaced and rotated. 268×252 pixel images are used in the experiments and the apples approximately account for 55% of the entire image; matching recognition time is 0.185 s without anticipation, and after anticipation, the time is 0.133 s, which means that the processing time of the improved algorithm is reduced by 28.1%. The comparison experiment demonstrates that the new method accelerates the speed of robot and makes it more practical.