33 / 2025-03-31 13:03:19

Toward Optimal Sampling Rate Selection for Precise Animal Activity Recognition

Precision Livestick Farming,animal behavior,wearable sensors

3. 智慧牧场与动物福利 > （2）智慧养殖的大数据、机器学习和物联网应用

Recent advancements in deep learning have significantly enhanced wearable sensor-based animal activity recognition, offering substantial improvements in livestock management efficiency and animal health monitoring. However, existing approaches often prioritize overall performance metrics, neglecting the uneven classification accuracy across specific behavioral categories—a limitation typically caused by suboptimal sampling rates. To address this challenge and ensure high-precision recognition for all individual behaviors in farm animals, we introduce a novel network. Specifically, given that distinct behaviors may require different sampling rates for optimal detection, the proposed network integrates multi-rate sensor data and employs a Mixture-of-Experts framework to softly fuse multi-scale features, thereby extracting behavior-specific representations. The proposed network’s efficacy is rigorously evaluated on two public datasets (goat and cattle behavior), with experimental results demonstrating its superior performance. Detailed findings will be presented in this paper.