Authors

Abstract

To address the impact of changes in dairy cow posture on intelligent body condition scoring, this study proposes a dairy cow body condition scoring model (ME-PointNeXt). Based on PointNeXt, this model incorporates a dynamic graph convolutional module at the global feature level; furthermore, a selective sequence block is added to the encoder’s upper layers, improving the SA-FP architecture and enhancing the model’s overall understanding of the structure. Using this model to segment the hindquarters of dairy cows, principal component analysis is performed via eigenvalue decomposition of the local neighbourhood covariance matrix. This enables the determination of edge point normal vectors and the calculation of surface curvature, thereby enabling the scoring of the cows’ body condition. Experimental results show that the ME-PointNeXt model achieved an overall accuracy, mean precision, mean recall, and mean F1 score of 91.43%, 84.33%, 87.92%, and 90.74%, respectively, on the test set. Compared with PointNeXt, the ME-PointNeXt model achieved improvements of 2.72, 3.65, 4.13, and 4.19 percentage points in overall accuracy, average precision, average recall, and average F1 score, respectively. This study has significantly improved the efficiency of dairy cow body condition scoring, ensuring accuracy whilst reducing the workload, and can provide technical support for non-contact dairy cow body condition scoring.

Abstract in Chinese

为解决奶牛姿态变化对体况智能评分的影响，本研究提出了一种奶牛体况评分模型（ME-PointNeXt）。该模型基于PointNeXt，在全局特征层加入一个动态图卷积模块；并且在编码器高层中加入一个选择性序列块，改进了SA-FP结构，增强整体对结构的理解。利用该模型对奶牛尻部进行分割，通过计算局部邻域协方差矩阵的特征值分解实现主成分分析，进而确定边缘点法向量并计算表面曲率，对奶牛体况进行评分。试验结果表明，ME-PointNeXt模型在测试集上的总体准确率、平均交互比、平均准确率与平均F1分数分别达到91.43%、84.33%、87.92%和90.74%。相较于PointNeXt，ME-PointNeXt的总体准确率、平均交互比、平均准确率与平均F1分数分别提升了2.72、3.65、4.13和4.19个百分点。本研究显著的提高了奶牛体况的评分效率，在降低工作量的同时保证了精度，可为非接触式奶牛体况评分提供技术支持。