Hardware platform and environment parameters.

Heatmap-based cattle pose estimation methods suffer from high network complexity and low detection speed. Addressing the issue of cattle pose estimation for complex scenarios without heatmaps, an end-to-end, lightweight cattle pose estimation network utilizing a reparameterized network and an attention mechanism is proposed to improve the overall network performance. The EfficientRepBiPAN (Efficient Representation Bi-Directional Progressive Attention Network) module, incorporated into the neck network, adeptly captures target features across various scales while also mitigating model redundancy. Moreover, a 3D parameterless SimAM (Similarity-based Attention Mechanism) attention mechanism is introduced into the backbone to capture richer directional and positional feature information. We constructed 6846 images to evaluate the performance of the model. The experimental results demonstrate that the proposed network outperforms the baseline method with a 4.3% increase in average accuracy at OKS = 0.5 on the test set. The proposed network reduces the number of floating-point computations by 1.0 G and the number of parameters by 0.16 M. Through comparative evaluations with heatmap and regression-based models such as HRNet, HigherHRNet, DEKR, DEKRv2, and YOLOv5-pose, our method improves AP0.5 by at least 0.4%, reduces the number of parameters by at least 0.4%, and decreases the amount of computation by at least 1.0 GFLOPs, achieving a harmonious balance between accuracy and efficiency. This method can serve as a theoretical reference for estimating cattle poses in various livestock industries.