【摘要】 在肉牛养殖中，肉牛质量对其生长监测和育种改良具有重要意义。传统的称量方法费时费力，且易造成牛只应激反应。然而，现有的非接触方法易受姿态和复杂背景影响，精度较低，鲁棒性差。为此，本文提出了基于改进YOLO v9u-pose的肉牛质量估算方法，包括关键点检测和质量估算2个阶段。在关键点检测阶段，以YOLO v9u-pose作为基线网络，利用ODConv(Omni-dimensional dynamic convolution)替换主干网络的普通卷积；采用DySample替换颈部网络的上采样模块；并在与检测头连接的RepNCSPELAN4模块中添加EMA注意力机制(Excitation and modulation attention),进而提高肉牛关键点检测算法精度。在质量估算阶段，利用深度图和局部点云聚类等方法提取体尺特征，并构建基于体尺和PSO-XGBoost(Particle swarm optimization-eXtreme gradient boosting, PSO-XGBoost)的肉牛质量估算算法。在自建的数据集上测试，本文提出的关键点检测算法F1值和平均精度均值(mAP@0.75)分别为97.2%和98.2%,质量估算算法平均绝对百分比误差为3.97%。最终将所提方法部署至开发板，为肉牛智能化养殖提供了技术支持。更多还原

【Abstract】 In beef cattle farming management, the weight of beef cattle is crucial for monitoring growth, improving breeding, and controlling costs. Traditional weighing methods are not only time-consuming and labor-intensive but also prone to causing stress in the cattle. However, existing non-contact estimation methods are easily affected by posture and complex backgrounds, leading to low algorithm accuracy and poor robustness. Therefore, a beef cattle weight estimation method was proposed based on improved YOLO v9u-pose, which mainly consisted of two stages: key point detection and weight estimation. In the key point detection stage, YOLO v9u-pose was used as the baseline model, where the standard convolutions in the backbone network were replaced with omni-dimensional dynamic convolution(ODConv); DySample was employed to replace the upsampling module of the neck network; additionally, the excitation and modulation attention(EMA) was added to the RepNCSPELAN4 module connected to the detection head to improve the accuracy of the beef cattle key point detection algorithm. In the weight estimation stage, body size parameters were extracted by using depth maps and local point cloud processing methods. A beef cattle weight estimation algorithm was then constructed based on the parameters and the particle swarm optimization-eXtreme gradient boosting(PSO-XGBoost) approach. On testing with a self-constructed dataset, the F1 score, and mean average precision(mAP@0.75) of the key point detection model proposed were 97.2% and 98.2%, respectively. The mean absolute percentage error(MAPE)of weight estimation method based on PSO-XGBoost was 3.97%. Finally, the proposed cattle weight estimation model was deployed to a development board, providing technical support for intelligent beef cattle farming.更多还原