张先洁, 汪小旵, 孙国祥, 等. 基于彩色点云图像的不同成熟阶段番茄果实数量的测定方法[J]. 华南农业大学学报, 2022, 43(2): 105-112. DOI: 10.7671/j.issn.1001-411X.202105021
    引用本文: 张先洁, 汪小旵, 孙国祥, 等. 基于彩色点云图像的不同成熟阶段番茄果实数量的测定方法[J]. 华南农业大学学报, 2022, 43(2): 105-112. DOI: 10.7671/j.issn.1001-411X.202105021
    ZHANG Xianjie, WANG Xiaochan, SUN Guoxiang, et al. Measurement of tomato fruits quantity at different ripening stages based on color point cloud images[J]. Journal of South China Agricultural University, 2022, 43(2): 105-112. DOI: 10.7671/j.issn.1001-411X.202105021
    Citation: ZHANG Xianjie, WANG Xiaochan, SUN Guoxiang, et al. Measurement of tomato fruits quantity at different ripening stages based on color point cloud images[J]. Journal of South China Agricultural University, 2022, 43(2): 105-112. DOI: 10.7671/j.issn.1001-411X.202105021

    基于彩色点云图像的不同成熟阶段番茄果实数量的测定方法

    Measurement of tomato fruits quantity at different ripening stages based on color point cloud images

    • 摘要:
      目的  为测定温室中番茄不同成熟阶段的果实数量,提出一种基于彩色点云图像的测定方法。
      方法  在移动平台上搭载KinectV2.0采集温室中行栽番茄的图像信息合成番茄植株点云,再将二视角的番茄植株点云合成1个点云,并通过深度信息截取得到近处番茄植株点云,将标注的点云数据输入到PointRCNN目标检测网络训练预测模型,并识别番茄植株点云中的番茄果实,最后利用基于特征矩阵训练的支持向量机(Support vector machine, SVM)分类器对已经识别出来的果实进行成熟阶段分类,获得不同成熟阶段番茄果实的数量。
      结果  基于PointRCNN目标检测网络的方法识别番茄果实数量的精确率为86.19%,召回率为83.39%;基于特征矩阵训练的SVM分类器,针对番茄果实成熟阶段的预测结果在训练集上准确率为94.27%,测试集上准确率为96.09%。
      结论  基于彩色点云图像的测定方法能够较为准确地识别不同成熟阶段的番茄果实,可以为评估温室番茄产量提供数据支撑。

       

      Abstract:
      Objective  In order to measure the number of tomato fruits at different ripening stages in greenhouse, a method based on color point cloud images was proposed.
      Method  The image information of tomato in greenhouse was collected by KinectV2.0 on the mobile platform to synthesize the tomato plant point cloud, then the tomato plant point clouds from two perspectives were synthesized into a point cloud, and the point cloud of nearby tomato plant was obtained by depth information interception. The labeled point cloud data were input into the PointRCNN object detection network to train the prediction model and recognize tomato fruit in the tomato plant point cloud. Finally, support vector machine(SVM) classifier based on feature matrix training was used to classify ripeness of the identified fruits, and the number of tomato fruits at different ripening stages was obtained.
      Result  The precision rate of the method based on PointRCNN object detection network for identifying the number of tomato fruits was 86.19% and the recall rate was 83.39%. The accuracy of SVM classifier based on feature matrix training for predicting the ripeness of tomato fruits was 94.27% in the training set and 96.09% in the test set.
      Conclusion  The measurement method based on color point cloud images can accurately identify the number of tomato fruits at different ripening stages, and provide data supports for evaluating the yield of tomato fruits in greenhouse.

       

    /

    返回文章
    返回