Abstract: To address the challenges in detecting phenotypic parameters such as diameter and erectness due to dense growth and complex occlusion of mature tomato seedlings (35-40 d) during the plug seedling stage, a dual-view detection method combining a variable-pitch stepped manipulator and the YOLO-SDCG model is proposed. The innovations of this work lie in three aspects: hardware design, improvements to the YOLOv8s-seg model, and phenotypic parameter extraction. First, in terms of hardware design, an image acquisition system for tomato seedlings was constructed, integrating top-view and side-view perspectives to capture image data. Second, regarding the vision algorithm, the YOLOv8s-seg model was enhanced to YOLO-SDCG by introducing dynamic snake convolution (DySConv) to focus on the slender tubular features of the main stem during the initial feature extraction stage. The content-aware reassembly of features (CARAFE) module was adopted to enhance the resolution of high-level semantic features of tomato seedling stems extracted from deep layers of the backbone network through content-adaptive upsampling. The resulting high-resolution stem feature maps were then concatenated or weighted-fused with high-resolution detailed features from shallow layers of the backbone network, achieving complementary integration of stem semantic information and seedling spatial details. Grouped hybrid one-shot tensor (GHOST) convolution was incorporated to extract intrinsic features using a small number of standard convolutions and generate complementary ghost features through linear transformation, thereby reducing model parameters and computational cost. Finally, for parameter extraction, based on stem segmentation results, image segmentation, elliptic Fourier descriptors (EFDs), the maximum inscribed circle method, chord-to-arc ratio, and piecewise fitting were integrated to detect phenotypic parameters such as stem diameter and erectness. Experimental results showed that in hardware design, the variable-pitch stepped manipulator enabled a stepped arrangement of multiple plants, effectively avoiding seedling occlusion from different perspectives, thereby providing a clear and stable input basis for subsequent image acquisition and model detection. In terms of the vision algorithm, compared to YOLOv8-seg, the YOLO-SDCG model achieved precision, recall, and mean average precision of 93.1%, 93.9%, and 94.9%, representing improvements of 4.6, 2.7, and 2.4 percentage points, respectively, with parameter count and inference time of 3.58 M and 3.0 ms, effectively segmenting the overall contour structure of stems while maintaining a good balance between accuracy and computational efficiency. For parameter extraction, the contour of the main stem of tomato seedlings is a closed tubular curve. Using elliptic Fourier descriptors with a harmonic order n=16 enabled high-quality reconstruction. The proposed maximum inscribed circle method yielded mean absolute errors of 0.03 mm for stem diameter in both top and side views, with mean absolute percentage errors of 1.04%. The chord-to-arc ratio and piecewise fitting method achieved mean absolute errors for erectness of 1.60° and 1.80° in top and side views, respectively, with mean absolute percentage errors of 2.00% and 2.14%, and coefficients of determination all exceeding 0.96. Transplanting experiments demonstrated that at a picking frequency of 120 seedlings per minute, the success rates for clamping seedlings from 72-cell and 105-cell trays were 93.13% and 92.50%, respectively, indicating that the transplanting system ensured operational efficiency while maintaining reliable seedling screening and grasping capabilities. This study has achieved preliminary results in tomato seedling phenotypic detection, though only two phenotypic traits—diameter and erectness—were included. Future work will expand the dataset to incorporate more phenotypic features and focus on model lightweighting to enable more comprehensive and efficient detection of tomato seedling phenotypes.