Tomato picking point detection and three-dimensional pose estimation based on keypoint detection and pedicel skeleton point cloud

Accurate acquisition of fruit position, picking points, and spatial orientation is essential for artificial-intelligence-enabled robotic harvesting of fresh-eating tomatoes. Existing systems often focus on fruit detection and picking-point localization, while three-dimensional orientation information, which is useful for collision-aware path planning, is less frequently integrated. This study develops a lightweight computer-vision framework based on You Only Look Once version 8 nano pose (YOLOv8n-Pose) for tomato detection and picking-point localization, together with a pedicel-skeleton-based three-dimensional pose estimation method. The detection network adopts StarNet as the backbone, a FasterNet-based module with a convolutional gated linear unit, and a lightweight shared-convolution detection head with detail-enhanced convolution. The improved model achieves a precision of 96.7%, while reducing the parameter count by 61.36% and the computational complexity to 4.2 giga floating-point operations. For pose estimation, the distal pedicel region is processed to extract a two-dimensional skeleton, which is combined with depth information to construct a three-dimensional skeleton point cloud. A spatial plane and circular arc are fitted, and the local tangent direction is used as the task-oriented tomato pose. The three-dimensional pose calculation requires approximately 52 ms after two-dimensional detection, with 93.4% of the evaluated samples exhibiting an angular error below 20°. The proposed framework reduces the amount of point-cloud data required for pose calculation and provides a computationally efficient perception approach for robotic tomato harvesting.