双目视觉自动检测香蕉植株假茎宽度

Automatic width detection of banana pseudostem with binocular vision

为了实现香蕉植株假茎宽度的快速无损测量,提出了一种基于双目相机的图像测量方法。实验随机选取100株香蕉植株分别从距香蕉假茎0.9、1.2、1.6、1.9 m处获取香蕉植株的双目RGB图像,共得到400个样本。首先通过双目标定、立体校正、立体匹配得到视差图,并将其随机分为训练集(170个样本)与测试集(230个样本)。然后对训练集中的假茎进行手动矩形框标注,并采用级联分类器进行训练,得到假茎的自动识别模型。最后将测试样本识别出的假茎矩形区域特征点转换到相机坐标系下进行倾斜度矫正,得到假茎的最终测量宽度。结果表明,本系统可以对香蕉假茎进行快速识别,识别率达到94.8%。不同的测量距离会影响检测精度,1.2 m距离处的测量精度最高,与真实值的决定系数R^(2)达到了0.8538,相对误差仅为2.0%。

In order to realize rapid and nondestructive measurement of the width of banana pseudostem,an image measurement method based on binocular camera was proposed.In the field experiments,100 banana plants were randomly selected to acquire their binocular RGB images at the distances of 0.9 m,1.2 m,1.6 m and 1.9 m respectively,collecting 400 pairs of image samples in total.Firstly,400 disparity maps were obtained in the process of binocular calibration,stereoscopic correction and stereoscopic matching,which were divided into training set(170 samples)and testing set(230 samples).Then the pseudostems in the training set were manually labeled with rectangular boxes,and were sent into cascade classifier to train a model which was used to automatically identify the locations of the pseudostems.Finally,the feature points of the identified rectangles of the pseudostem in the testing samples were converted to the camera coordinate system for inclination correction,and the final stem width was calculated.The results show that the system could quickly identify and locate the banana pseudostem with an accuracy of 94.8%.Different measurement distances will affect the width detection accuracy,the measurement accuracy at the distance of 1.2 m is the highest among all the distances,reaching R^(2) of 0.8538,with a relative error of only 2.0%.