Cattle Behaviour Recognition from Multi-IMU Collar Data with a Calibration-Free Descriptor Network: Cross-Animal Evaluation and Microcontroller Feasibility
Collar-mounted inertial sensors enable continuous cattle behaviour monitoring, but classifiers are usually evaluated with random window-level splits that overstate performance on unseen animals. Using a public dataset of ten free-grazing Holstein cows (four behaviours), we asked which model family generalises best to unseen animals, using leave-one-cow-out validation and a fully inductive pipeline. No end-to-end temporal encoder (CNN, CNN-Transformer, four hybrid fusion variants, four training regimes) matched a small network on window statistics. Extending 168 generic statistics with 16 cattle-specific descriptors (dual-IMU agreement, head posture, gait rhythm) gave a calibration-free network (6420 parameters) reaching 79.1 ± 0.6% accuracy and 74.5 ± 0.6% macro-F1, versus 69.6% and 67.3% for a Random Forest; with ten animals, the per-cow difference was not significant, and a selection-aware nested estimate was lower (72%). Gyroscope and second-IMU data improved recognition. Per-animal standardisation over the test recording proved transductive and failed when estimated from a calibration period (52% accuracy). The INT8 C implementation needed 8.9 kB of flash and about 150 k cycles per window on an emulated Cortex-M4F, indicating feasibility, not validated deployment. Federated training reached 74.0% on unseen cows, 1.7 points below centralised training with identical passes. Resting–standing confusion and the single-farm dataset limit practical use.
Authors
- Subhash Jayakumar
- Xavier Arockiaraj Santhappan (ORCID: https://orcid.org/0000-0002-6420-7364)
Publication Details
- Journal
- Sensors
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3390/s26227272
- Primary Topic
- Smart Agriculture and AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00