Study of Pose Estimation Based on Spatio-Temporal Characteristics of Cow Skeleton

Abstract

The pose of cows reflects their body condition, and the information contained in the skeleton can provide data support for lameness, estrus, milk yield, and contraction behavior detection. This paper presents an algorithm for automatically detecting the condition of cows in a real farm environment based on skeleton spatio-temporal features. The cow skeleton is obtained by matching Partial Confidence Maps (PCMs) and Partial Affinity Fields (PAFs). The effectiveness of skeleton extraction was validated by testing 780 images for three different poses (standing, walking, and lying). The results indicate that the Average Precision of Keypoints (APK) for the pelvis is highest in the standing and lying poses, achieving 89.52% and 90.13%, respectively. For walking, the highest APK for the legs was 88.52%, while the back APK was the lowest across all poses. To estimate the pose, a Multi-Scale Temporal Convolutional Network (MS-TCN) was constructed, and comparative experiments were conducted to compare different attention mechanisms and activation functions. Among the tested models, the CMS-TCN with Coord Attention and Gaussian Error Linear Unit (GELU) activation functions achieved precision, recall, and F1 scores of 94.71%, 86.99%, and 90.69%, respectively. This method demonstrates a relatively high detection rate, making it a valuable reference for animal pose estimation in precision livestock farming.

Keywords

• cows
• skeletons
• pose estimation
• attention mechanisms