Abstract: Pod count and seed count per pod are two of the most important agronomic traits to accurately estimate soybean yield. These phenotypic parameters can also be used to screen superior and high-yielding varieties, such as genetic traits. However, in-situ measurement is limited by the complex background interference from soil and foliage, the severe inter-pod occlusions caused by dense planting, and the high morphological similarity among pods with varying numbers of seeds under the field environment. Particularly, conventional manual phenotyping cannot fully meet the needs of large-scale cultivation in recent years, due to high labor intensity, time-consuming nature, and subjective errors. Consequently, high accuracy, efficiency, and robustness are often required for pod detection. In this study, a fully robust and multi-stage framework was developed and then validated for the classification, detection, and precise counting of soybean pods at the individual plant level in a natural field. 1) An advanced model of object detection, named TE-YOLO, was constructed to achieve the precise directional capture of the individual soybean plants. This approach mitigated the severe visual interference and overlapping from the adjacent plants under dense planting environments. 2) A five-point prompting strategy was proposed for inclined morphology of plants using geometric priors of rotated bounding boxes. Attention was anchored into the main axis of the plant, thereby driving the segment anything model (SAM) to perform high-quality instance segmentation of targeted plants. 3) An enhanced post-processing algorithm was integrated into the overall pipeline. High-quality instance segmentation of individual soybean plants was achieved to remove the complex background noise and irrelevant clutter found in the field. 4) A soybean pod counting network, designated as SoyPod-FGC, was constructed specifically for the purified images after segmentation. Feature extraction was then strengthened to reconstruct content awareness. Consequently, the edges and boundaries of pods were enhanced under dense occlusion. This robust architecture enabled accurate classification and identification of pods with the number of seeds. A series of experiments were conducted on a custom-built field dataset to validate the framework. The experimental results demonstrated that the multi-stage strategy overcame the adverse impacts caused by the complex backgrounds and the severe inter-pod occlusions. Specifically, the mean average precision (mAP) of the SoyPod-FGC model reached an impressive 74.4%. The positive superior performance achieved a 1.5 percentage point improvement, compared with the baseline YOLO11m model. Furthermore, the mean absolute error (MAE) was reduced to 3.50 for the pod counting per individual plant. Concurrently, the coefficient of determination for the seed counting reached a reliable 0.816. In conclusion, the multi-stage model can elevate the counting accuracy and anti-interference stability, particularly when processing densely overlapping targets. The accurate, robust, and practical approach can be deployed for real-world agricultural monitoring. The efficient pipeline can provide solid phenotypic acquisition support for continuous, high-throughput digital breeding, precise yield estimation, and data-driven crop of soybeans in precision agriculture and crop cultivation.