基于双目立体视觉的机械手移栽穴盘定位方法

为了解决自动化机械手移栽过程中穴盘放置偏斜和底部局部"凸起"而引起的移栽效果不理想的问题,为机械手提供穴盘精准坐标,对穴盘准确定位方法进行研究。首先,根据机械手移栽特点提出穴盘定位总体方法及图像获取手段。其次,利用单目相机获取的图像采用像素标记法和Radon变换法计算穴盘中心坐标和角度,完成穴盘平面定位。再次,对双目相机获取的图像采用SIFT(scale invariant feature transform)特征匹配的算法获得匹配点对坐标,并提出区域整合匹配点的方法。最后,利用整合的区域双目匹配点坐标配合相机标定结果重建匹配点的三维世界坐标,并且与穴盘平面定位结果相结合完成穴盘空间位置重构。试验结果表明,提出的穴盘定位方法能够真实地恢复穴盘空间姿态,中心像素横纵坐标相对误差分别在(-7,+7)和(-6,+7)像素内,角度检测值与实测值相对误差值在(-0.51°,+0.53°)内,利用SIFT特征匹配算法匹配双目图像,在2×4区域内对8对整合匹配点进行三维世界坐标重建,其中7个坐标的三个维度与测量值相对误差在2 mm内,1个坐标与测量值相对误差为4.6 mm内。该方法所应用的算法成熟,可以满足机械手移栽实际应用处理要求。

Mechanical transplanting plug tray localization method based on binocular stereo vision

Considering mechanical vibration and other external factors, there are two main aspects influence the plug tray position in the process of mechanical transplanting seedlings from plug tray to nutrition bowl. One is that the plug tray is skew or the center point of it has a displacement when the plug tray was placed on the manipulation platform or the conveyor belt area. The other is that the plug tray is raised partly which attributes to the soil adhesion at the bottom of it or the soil of previous plug tray accidentally slip onto the platform. Therefore, transplanting error can appear or fingers inserted too deep to wear out plug tray when manipulator walks according to the standard coordinates. These affect the transplanting quality, and sometimes even lead to failure of the whole plug tray transplant. In order to solve the problems and to provide precision coordinates of plug tray, we studied plug tray accurate localization method using binocular stereo vision. Firstly, we proposed the overall method of mechanical transplanting plug tray localization based on binocular stereo vision and acquisition of binocular images. Second, we located two-dimensional plane plug tray by calculating center point coordinates and deviation angles using images obtain from monocular camera. For calculations of center point, we used pixel labeled algorithm based on recursive method of connecting area in binary images, and angles adopted Radon transformation in edge detection images. Then, we located three-dimensional stereo plug tray by acquiring the coordinates of pairs of matching points that adopted Scale Invariant Feature Transform (SIFT) feature matching algorithm using pairs of images obtained from binocular cameras. Meanwhile, considering the factors of plug tray material and improvement the speed of three-dimensional reconstruction, we proposed a method of regional integrated matching points. Finally, we adopted Zhang camera calibration method, and used 8×8 checkerboards to calibrate stereo cameras. Cooperated with binocular camera calibration results, we rebuilt the three-dimensional coordinates of regional integrated matching points, and reconstructed plug tray spatial location by combining the results of two and three dimensions. Experimental results indicated that our plug tray localization method based on binocular stereo vision was able to reconstruct plug tray space posture veritably. We randomly selected 10 plug trays to do two-dimensional plane localization experiment. From the perspective of angle, we observed that the angle relative errors between detected and measured were in the range of -0.51° to +0.53°, in which the small angle had a large percentage of relative error. The reason for that was the Radon transform results were integer. From the perspective of center point coordinates, we observed the relative errors of horizontal and vertical coordinates between detected and measured were in the range of -7 to +7 and -6 to +7 pixels, respectively. The main source of error was the method taking the two-dimensional projection problem into account only. Angle and center point coordinates errors were within the acceptable range. We selected No.1 plug tray to do three-dimensional stereo localization experiment. We adopted SIFT feature matching algorithm matching binocular images (images size are 736 pixels×480 pixels),and got 147 pairs of matching points. When optimize the matching points by using RANSAC(RANdom SAmple Consensus) method, the number of matching points pairs dropped to 55. We reconstructed eight pairs of integrated matching points in the 2×4 area of world coordinates, seven of which the three dimensions relative errors between detected and measured were in 2 mm, and one of which was in 4.6 mm. In summary, our application algorithms were mature, the whole method was reliable, and satisfied the requirements of practical mechanical transplanting application processing.