基于双目视觉的葡萄识别定位及跟踪方法研究

Positioning and Tracking Method Based on Binocular Vision

为消除不同光照强度和葡萄叶片遮挡等因素对果穗识别定位与跟踪精确度的影响,通过搭建基于vs2017与OpenCV3.4双目识别定位系统,研究基于自然光环境下红提葡萄果穗的识别、定位及跟踪,提出一种K-Means聚类算法和HSV颜色分量相结合的葡萄果穗识别分割方法,并提出了基于深度卷积神经网络与双目视觉相结合的葡萄跟踪算法。选用OpenCv与MATLAB对双目相机标定,去除误差过大及模糊的照片,标定误差控制在0.13%以内。采用双目相机视差获得深度图进行葡萄定位,试验表明,当定位距离在0.8~1.5 m时,葡萄定位误差小于1%;利用轻量级卷积神经网络YoLov4训练葡萄位置区域模型,试验结果表明葡萄位置区域检测平均准确率达85.4%,召回率达87.6%;利用MATLAB对葡萄轮廓模型实施匹配跟踪和kalman滤波,试验结果表明跟踪效果较好。

In order to eliminate the influence of factors such as different light intensity and occlusion of grape leaves on the accuracy of ear recognition,positioning and tracking,a grape ear recognition and segmentation method combining K-Means clustering algorithm and HSV color component was put forward by establishing a binocular recognition and positioning system based on vs 2017 and OpenCv 3.4 to study the recognition and location of red grape ears under natural light environment and a grape tracking algorithm based on the combination of deep convolutional neural network and binocular vision was proposed as well.At the same time,OpenCV and MATLAB were selected to calibrate the binocular camera to remove too large error and blurry photos,with calibration errors controlled within 0.13%.Binocular camera parallax used to obtain the depth map for grape positioning,the experiment shows that where the positioning distance was 0.8-1.5 m,the grape positioning error was less than 1%;the lightweight convolutional neural network YoLov4 used to train the grape location area model,the experimental results show the average accuracy rate of grape location area detection was 85.4%,and the recall rate was 87.6%;the grape contour matching and tracking model extracted from MATLAB,and Kalman filtering applied to improve tracking accuracy,the experimental results show that the tracking enhanced effect was better.