基于关键点检测的动态山羊体尺测量

Dynamic Body Size Measurement of Individual Goats Based on Keypoint Detection

摘要:

山羊体尺测量是养殖管理的基础性环节，可提供量化生产性能、解析遗传特性、提升选种配准确性，监控健康状态的核心指标，对提升养殖经济效益与优化遗传育种方案发挥着不可替代的决策支撑作用。针对多姿态下动态山羊体尺测量困难问题，本研究提出了基于山羊关键点的多姿态体尺测量方法。采用改进YOLO v8n-Pose深度学习模型YOLO v8n-SK对通道内山羊进行关键点检测，该模型特别侧重于对山羊膝关节关键点的精确检测，旨在动态运动状态下最大限度地降低运动对体高测量的干扰：引入空间与通道重建卷积（ScConv）模块优化卷积到特征（C2f）结构，减少参数冗余的同时，增强模型的特征提取能力;加入CBAM注意力机制模块提高模型的抗干扰能力;并采用EIoU损失函数进一步优化模型训练过程，提升检测框的定位精度。进一步地从提取到的关键点坐标中获得山羊体尺相关特征，并结合单目深度估计数据，通过非线性回归模型对山羊体尺进行预测。实验结果表明，该模型在山羊躯干及背部关键点检测方面展现出优异的性能：检测精确率和召回率分别达98.0%和98.1%，mAP@50-95提升至95.8%，模型复杂度显著降低，参数量和浮点计算量分别为6.7×10?和7.9×10?。体尺参数预测中，体高、胸围、腹围及体斜长平均绝对百分比误差分别为2.408%、1.731%、1.340%和2.519%，表现出较高的测量精度。与传统人工测量方法相比，本方法能够大幅提升山羊体尺测量效率，并避免山羊受到应激刺激，在不干扰山羊情况下较为准确、快速地预测山羊体尺。

Abstract:

Body measurement of goats is a fundamental aspect of livestock management, serving as a core indicator for quantifying production performance, analyzing genetic traits, improving the accuracy of breeding selection, and monitoring health status. It plays an irreplaceable role in enhancing economic efficiency and optimizing genetic breeding strategies. To address the challenges of measuring goat body dimensions under dynamic and multi-posture conditions, a non-contact, rapid, and accurate measurement method was proposed. An improved YOLO v8n-Pose deep learning model, named YOLO v8n-SK, was developed for keypoint detection of goats in the channel environment. This model specifically focused on accurately detecting the knee joints of goats to minimize the interference of movement on height measurements during motion. The model introduced a spatial-channel collaborative reconstruction convolution (ScConv) module to optimize the convolution-to-feature (C2f) structure, reducing parameter redundancy while enhancing feature extraction capability. CBAM attention mechanism was integrated to improve the model's robustness to interference, and an EIoU loss function was adopted to further optimize the training process and improve bounding box localization accuracy. From the extracted keypoints, body measurement-related features were obtained and combined with monocular depth estimation data. A nonlinear regression model was then employed to predict the goat body dimensions. Experimental results demonstrated that the proposed model exhibited excellent performance in detecting trunk and back keypoints, achieving a precision of 98.0%, recall of 98.1%, and mAP@50-95 of 95.8%. Meanwhile, the model complexity was significantly reduced, with parameters and FLOPs controlled at 6.7×10? and 7.9×10?, respectively. In the prediction of body measurement parameters, the mean absolute percentage errors (MAPE) for height, chest girth, abdominal girth, and body diagonal length were 2.408%, 1.731%, 1.340% and 2.519% respectively, demonstrating high measurement accuracy. Compared with traditional manual methods, the proposed approach significantly improved measurement efficiency and avoided stress responses in goats, enabling fast and accurate body measurement without disturbing the animals.