基于STPM策略的肉牛出栏评估方法

Cattle Slaughter Evaluation Method Based on STPM Strategy

摘要:

在我国小型养殖场分散管理的背景下，肉牛出栏作为肉牛养殖的关键环节，高效且准确地掌握出栏时机是提高养殖效益的重点。针对该问题，本研究提出了一种基于实例分割与多目标跟踪的STPM（Segment-Track-PixelMax）智能出栏判定策略。首先，构建改进的GDL-YOLO 11n-seg模型，以实现对肉牛躯体轮廓的高精度实例分割，生成边界清晰的掩膜区域;随后，结合OC-SORT算法为每头肉牛分配唯一身份ID，实现持续跟踪，在时间序列上动态记录个体像素面积的变化;最后，提取全周期内每头肉牛的最大像素面积，并基于设定的阈值完成肉牛出栏状态判定。试验结果表明，该研究具有较高的分割精度与良好的跟踪效果，GDL-YOLO 11n-seg模型平均精度均值为94.4%，比原模型提高2.1个百分点，参数量和浮点运算量分别减少27.3%和24.5%;OC-SORT算法的MOTA、MOTP、HOTA和IDF1分别为87.7%、83.3%、79.3%和93.6%。研究结果表明，本文策略能够在真实牛圈环境中实现对肉牛出栏判定，为肉牛智能化养殖提供有效的技术支持。

Abstract:

In the context of decentralized management in small-scale livestock farms in China, efficient and accurate control of cattle slaughter timing is a critical factor for improving farming profitability. To address this issue, an intelligent slaughter decision strategy was proposed, Segment-Track-PixelMax (STPM), based on instance segmentation and multi-object tracking. Firstly, an improved GDL-YOLO 11n-seg model was developed to achieve high-precision instance segmentation of cattle body contours, generating clear boundary masks. Nextly, the OC-SORT algorithm was combined to assign a unique identity ID to each cow, enabling continuous tracking and dynamically recording the individual pixel area changes in the time series. Finally, the maximum pixel area of each cow over the entire period was extracted, and slaughter status was determined based on a set threshold. Experimental results demonstrated that the proposed method achieved good segmentation accuracy and tracking performance. The segmentation accuracy of the GDL-YOLO 11n-seg model was 94.4%, which represented an improvement of 2.1 percentage points over the original model, with reductions of 27.3% and 24.5% in parameter count and floating-point operations, respectively. The MOTA, MOTP, HOTA, and IDF1 of the OC-SORT algorithm were 87.7%, 83.3%, 79.3%, and 93.6%, respectively. The research result showed that this strategy can effectively perform cattle slaughter decision-making in real-world barn environments, providing strong technical support for intelligent cattle farming.