Geometry-Aware 3D Point Cloud Learning for Precise Cutting-Point Detection in Unstructured Field Environments

Abstract

In automated lychee harvesting, the complex geometric structures of branches, leaves, and clustered fruits pose significant challenges for robotic cutting point detection, where even minor positioning errors can lead to harvest damage and operational failures. This study introduces the Fcaf3d-lychee network model, specifically designed for precise lychee picking point localization. The data acquisition system utilizes Microsoft's Azure Kinect DK time-of-flight camera to capture point cloud data through multi-view stitching, enabling comprehensive spatial information capture. The proposed model enhances the Fully Convolutional Anchor-Free 3D Object Detection (Fcaf3d) architecture by incorporating a squeeze-and-excitation (SE) module, which leverages human visual attention mechanisms to improve feature extraction capabilities. Experimental results demonstrate the model's superior performance, achieving an $F_1$ score of 88.57% on the test data set, significantly outperforming existing approaches. Field tests in real orchard environments show robust performance under varying occlusion conditions, with detection accuracies of 0.932, 0.824, and 0.765 for unobstructed, partially obstructed, and severely obstructed scenarios, respectively. The model maintains localization errors within $\pm$1.5 cm in all directions, demonstrating exceptional precision for practical harvesting applications. This research advances the field of automated fruit harvesting by providing a reliable solution for accurate picking point detection, contributing to the development of more efficient agricultural robotics systems.