Abstract:In complex greenhouse environments, cherry tomatoes are easily affected by dense fruit occlusion, slender pedicels, and variable fruit postures, which makes pickingpoint localization difficult. To address this problem, a tomato small-target picking-point localization method based on the Sim-YOLO-Pose algorithm and a far-to-near multi-stage localization strategy is proposed. First, a Sim-YOLO-Pose model is constructed by integrating the parameter-free SimAM attention mechanism into the YOLOv5 backbone network and combining it with the YOLOv7-pose model, thereby enhancing the detection capability for small tomato targets and pedicels while suppressing complex background interference. Second, a far-to-near multi-stage localization process is designed. At a long distance, the Sim-YOLO model is used to detect tomato maturity and screen mature fruits. The camera is then dynamically adjusted to a close distance, where YOLOv7-pose is used to detect fruit contour keypoints and estimate spatial posture. Subsequently, the YOLACT instance segmentation model is used to extract the pedicel mask, and the pedicel position is predicted by combining posture information. Finally, the pedicel contour and fruit keypoint coordinates are fused to obtain the three-dimensional coordinates of the picking point through geometric calculation. Experimental results show that the Sim-YOLO-Pose model achieves an mAP@0.5 of 97.5% in the detection of small tomato targets and pedicels, which is 4.5 percentage points higher than that of YOLOv5-pose, with a single-frame detection time of 0.022 s. The far-to-near multi-stage localization method achieves a picking-point recognition success rate of 86.7% in complex scenes, and the average detection confidence of single-fruit keypoints in close-range images reaches 51%. The results indicate that the proposed method can effectively improve the accuracy and robustness of small-target detection and picking-point localization for cherry tomatoes in complex greenhouse environments, meeting the requirements of real-time harvesting applications.