Abstract: Phenotypic parameters can be detected, such as the diameter and erectness, under dense growth and complex occlusion at the mature tomato seedlings (35-40 d) during the plug seedling stage. In this study, a dual-view detection was proposed to combine a variable-pitch stepped manipulator and the YOLO-SDCG model. Three aspects were employed, including the hardware design, the improved YOLOv8s-seg model, and phenotypic parameter extraction. Firstly, in terms of the hardware design, an image acquisition was constructed for the tomato seedlings. The top- and side-view perspectives were integrated to capture the image data. Secondly, the YOLOv8s-seg model was enhanced into the YOLO-SDCG model, in terms of the vision algorithm. Dynamic snake convolution (DySConv) was introduced to extract the slender tubular features of the main stem during the initial feature extraction. The content-aware reassembly of the features (CARAFE) module was adopted to enhance the resolution for the high-level semantic features. The tomato seedling stems were extracted from the deep layers of the backbone network using content-adaptive upsampling. The high-resolution stem feature maps were then concatenated or weighted-fused from the shallow layers of the backbone network. Stem semantic information was integrated with the seedling spatial details. Grouped hybrid one-shot tensor (GHOST) convolution was incorporated to extract the intrinsic features using a small number of standard convolutions. The ghost features were generated after a linear transformation. Thereby, the parameters and computational cost were reduced for the parameter extraction. Finally, the image segmentation, elliptic Fourier descriptors (EFDs), the maximum inscribed circle, chord-to-arc ratio, and piecewise fitting were integrated to detect the phenotypic parameters, such as the stem diameter and erectness, after stem segmentation. Experimental results showed that a stepped arrangement of the multiple plants was achieved to effectively avoid the seedling occlusion using the variable-pitch stepped manipulator. Thereby, a stable input was provided for the subsequent image acquisition and detection after hardware design. In terms of the vision algorithm, compared with the YOLOv8-seg, the YOLO-SDCG model was achieved in the precision, recall, and mean average precision of 93.1%, 93.9%, and 94.9%, respectively, which was improved by 4.6, 2.7, and 2.4 percentage points, respectively, with the parameter count and inference time of 3.58 M and 3.0 ms, respectively. The contour structure of the stems was effectively segmented to maintain a better balance between accuracy and computational efficiency. The contour of the main stem of the tomato seedlings was a closed tubular curve after parameter extraction. The high-quality reconstruction was obtained using elliptic Fourier descriptors with a harmonic order n=16. The maximum inscribed circle yielded the mean absolute errors of 0.03 mm for the stem diameter in both top and side views, with the mean absolute percentage errors of 1.04%. The chord-to-arc ratio and piecewise fitting were achieved in the mean absolute errors for the erectness of 1.60° and 1.80° in the top and side views, respectively, with the mean absolute percentage errors of 2.00% and 2.14%, while the coefficients of determination all exceeded 0.96. Transplanting experiments demonstrated that the success rates were 93.13% and 92.50%, respectively, for the clamping of seedlings from 72-cell and 105-cell trays at a picking frequency of 120 seedlings per minute. The transplanting shared the operational efficiency to maintain the reliable seedling screening and grasping. Only two phenotypic traits—the diameter and erectness—were included in the phenotypic detection of the tomato seedling. The dataset can be expected to incorporate more phenotypic features in the more efficient detection of the tomato seedling phenotypes.