Design and implementation of sheep posture estimation algorithm based on skeletal keypoint localization

Abstract

Automated sheep posture recognition can provide posture-related behavioural information for precision livestock monitoring. Manual video observation is labour-intensive, and contact sensor-based methods may require specialized installation. This study developed a non-contact sheep posture recognition method by combining skeletal keypoint localization with angular feature-based classification. A sheep posture dataset, SS-10K, was constructed using 10,000 images from five posture categories and annotated with 11 predefined skeletal keypoints. A ResNet-50-based network was adapted for sheep skeletal keypoint localization. Angular features were calculated from the localized keypoint coordinates. The complete 111-dimensional angular representation was used in the main posture-classification experiment, while ANOVA-based feature ranking was separately evaluated to obtain a reduced top-10 representation. K-nearest neighbours, support vector machine, and backpropagation neural network classifiers were compared for sheep posture classification. The ResNet-50-based keypoint localization method achieved mean average precision (mAP) values of 0.825, 0.884, 0.840, 0.881, and 0.831 for forelimb bending, grazing, lateral lying, standing, and kneeling on the forelimbs, respectively. Using the 111-dimensional angular representation, the BP neural network optimized by the Conjugate Gradient with Powell/Beale Restarts algorithm achieved a mean posture classification accuracy of 91.76 ± 0.23% across five repeated runs under the stratified image-level evaluation protocol. The results demonstrate the feasibility of combining skeletal keypoint localization with angular feature-based classification for non-contact sheep posture recognition under the evaluated SS-10K conditions. Further validation under more diverse farming and imaging conditions is required to assess the broader applicability of the proposed framework.