Advancing Precision Livestock Farming: Robust Country Chicken Detection via FeatherNet Fusion-YOLO and HenSense

Maneesh Reddy Bhavanam, Pavan Suraj Kanna, Tarun Sai Reddy Tamma, Chaitanya Alapati; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2026, pp. 530-538

Detecting free-range poultry in unconstrained farm environments remains challenging due to appearance variation, occlusion, illumination changes, and background complexity. We present HenSense, a large-scale dataset for robust country chicken detection comprising 13,995 high resolution images with 195,163 bounding box annotations collected from multiple farms across diverse viewpoints, lighting conditions (day/night/mixed), occlusion levels, and densities. The dataset includes rich metadata for occlusion severity, lighting category, age group, and annotation provenance, enabling fine-grained diagnostic evaluation. A semi-automated LangSAM-guided annotation pipeline with human verification reduces labeling effort by 80% while maintaining high annotation quality. We benchmark YOLOv5/8/9/10/11/12 and propose FeatherNet Fusion-YOLO, our lightweight attention-enhanced detector achieving 92.6% mAP@0.5 and 72.3% mAP@0.5:0.95 with only 2.14M parameters--suitable for edge deployment. Performance is analyzed across occlusion levels, lighting conditions, and collection sites. HenSense advances agricultural vision through the first comprehensive benchmark for native poultry detection under realistic smallholder farm conditions.