Answers are grounded in this paper and cite their sources.
Dairy goats are important livestock animals, and their daily behaviors are closely related to individual health, welfare, and milk production performance. However, the accurate recognition and localization of their behaviors in complex indoor environments remains a challenge in intelligent livestock farming. To address this, here we present a comprehensive framework covering both dataset construction and model design. Considering the high annotation cost, we propose a semi-automatic annotation pipeline and construct a benchmark for multi-label spatio-temporal goat behavior detection in surveillance videos, termed GoatBehav , which provides a solid foundation for model training and evaluation. To mitigate the limitation of fixed fusion between 2D spatial features and 3D temporal features, we further present the S patio- T emporal G oat B ehavior D etection framework, termed STGBD, which contains two novel blocks. The first is Decoupled Adaptive Local Attention (DALA) module, which adaptively integrates motion information and enhances motion-sensitive channels for both classification and regression. The second is Instance-level Feature Refinement (IFR) block that leverages instance specific geometric cues from the regression branch to refine classification representations. Our STGBD is designed to be lightweight, achieving accuracy on par with mainstream models on GoatBehav, a dense livestock farming benchmark, with substantially lower computational cost. We also perform extensive ablations to analyze the effectiveness of our model design, and disclose its shortcomings for future investigation.
No takes yet. Share an insight, caveat, or question.
Yi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: