3D point cloud-based 6D pose estimation using the pedicel morphological features of fruits and vegetables

This study proposes a method to estimate the 6D pose of pedicels using the morphological features of fruits and vegetables. The pedicel, a critical element connecting fruits to their stems, significantly influences the precision and efficiency of agricultural harvesting robots. The proposed system employs 3D point cloud data obtained from RGB-D cameras, using differences in width and curvature to identify the location and orientation of the pedicel. A lightweight YOLOv8n-seg architecture is employed to detect fruits and vegetables, computing the local curvature to estimate pedicel positions. The experimental evaluations on tomatoes and Cucumis melo (C. melo) demonstrate the capability of the proposed system to handle diverse agricultural environments. For C. melo, the system achieved a precision of 0.927, recall of 0.809 and F1-score of 0.864. For tomatoes, precision and recall were both 0.837, resulting in an F1-score of 0.837. Positional errors along the $x$-, $y$- and $z$-axes averaged 1.34, 3.19 and 4.79 mm , respectively, for the C. melo, with corresponding root mean squared errors of 7.95, 5.46 and 5.13 mm. Orientational errors averaged 2.14°, 1.14°and -1.49°for $\varphi$, $\theta$ and $\psi$, respectively. Smoothing algorithms, including linear interpolation for translation and spherical linear interpolation for rotation, address positional and orientational instability, further enhancing trajectory precision. The system achieved real-time operation with a processing speed exceeding 20 fps with smoothing, making it suitable for dynamic agricultural tasks. The results highlight the robust performance of the system in accurately identifying and approaching pedicels, even in occluded or clustered conditions.