Proposal for Method of Estimating Fruit Volume of Cherry Tomatoes Using 3D Image Measurement

Abstract

In this study, an ortho image was generated from the 3D point cloud data of cherry tomatoes, and fruit regions were extracted using SegNet, a deep learning based semantic segmentation model. From the extracted fruit regions, a reconstructed 3D point cloud of each fruit was generated, and the occluded regions were complemented based on the fruit's geometric symmetry. The fruit volume was then estimated using the least-squares ellipsoidal approximation.

When the fruit was not occluded by leaves, the estimation error was within ±1 ml, demonstrating the effectiveness of the proposed method. In contrast, when the fruit was partially occluded by leaves, the error increased significantly to 18.66±14.33 ml, due to incomplete reconstruction and poor shape fitting with an ellipsoid. To address this issue, a novel method was introduced to estimate the fruit's center and symmetrically complement occluded regions. As a result, the accuracy of the volume estimation improved substantially, reducing the error to 1.73±3.91 ml.