Abstract: In order to realize the online 3D size estimation of strawberries grown on monopolies under natural growing conditions, this paper proposed a lightweight framework based on the improved YOLOv8-seg and semi-truncated cone model. By integrating the C2f_Faster_EMA module into the YOLOv8-seg backbone, the network elevated mAP@0.5 from 94.1% to 96.9% and achieved a ripe-fruit segmentation accuracy of 98.3% while retaining an inference speed of 159 FPS, at the marginal cost of merely 0.3M additional parameters and less than 3% increase in model size. Secondly, the point cloud data were acquired by RealSense D435i, and the main axis and the centroid estimation were completed by ‘point cloud matching-statistical filtering-voxel down-sampling-RANSAC-PCA cascade’ processing. Then, a semi-truncated cone model was constructed with three parameters, namely, longitudinal diameter, top diameter and bottom diameter. Sub-millimetre alignment between the model and the observed point cloud was achieved through L-BFGS-B optimization constrained by a bidirectional chamfer distance. Field experiments demonstrated that, across different capture angles, the relative deviations between estimated and measured longitudinal, top, and bottom diameters were 1.2%, 0.7%, and 1.7%, with standard deviations of 0.64 mm, 0.34 mm, and 0.17 mm, respectively. Under different capture heights, the corresponding deviations were 1.8%, 1.2%, and 4.8%, all exhibiting standard deviations were below 0.6 mm. In multi-sample experiment, the mean estimation error ranges from 0.22% to 1.34%, with standard deviations of 2.16 mm, 1.67 mm, and 1.40 mm, and maximum ranges remaining under 2 mm. The complete processing pipeline averaged 3.3s per frame, fulfilling the real-time requirements for in-field dimensional estimation. Accordingly, the proposed methodology delivered a transferable, resource-efficient, and high-accuracy paradigm for three-dimensional phenotypic characterization of conical fruit and vegetable crops.