Cattle posture recognition in complex farm environments with real-time edge deployment
Cattle posture acts as a core measurable indicator for cattle health monitoring and animal welfare evaluation. Nevertheless, accurate and reliable posture recognition under the evaluated farm conditions is hindered by diverse shooting scenes, messy background interference, frequent target occlusion, and drastic object scale changes. Existing deep learning detection methods have achieved preliminary progress, yet two critical challenges remain in practical deployment: insufficient robustness under scene variations and incomplete real-time verification on low-power edge hardware. This work designs a lightweight detection framework Cattle-YOLO for multi-scenario farm cattle posture recognition. We propose C3k2-DWRSeg for multi-receptive-field structural feature extraction; a lightweight MLCA attention block suppresses background noise; and an SDI fusion module mitigates cross-scale semantic mismatch. Experiments are conducted on a self-built multi-source dataset covering four farm scenarios. Cattle-YOLO achieves [email protected] of 95.8%, outperforming YOLO11n by 3.1%. After TensorRT FP16/INT8 quantization on Jetson Orin Super, the average single-frame latency reaches 7.5 ± 0.4 ms, verifying stable long-duration real-time monitoring capacity.
Authors
- Bing Yang (ORCID: https://orcid.org/0000-0002-0585-0579)
- Haibo Pu (ORCID: https://orcid.org/0000-0001-6344-4053)
- Haohui Peng
- Xiaofan Li
- shangyang Ju
- Zhongbin Yan
Institutions
- Sichuan Agricultural University (CN)
Publication Details
- Journal
- PLoS ONE
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1371/journal.pone.0357974
- Primary Topic
- Smart Agriculture and AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00