A Lightweight Deep Learning Model for Broiler Population Monitoring on an Edge AI Platform

Although lightweight deep learning models have shown promise for livestock monitoring, there is still limited evidence regarding their comparative performance and practical deployment under real broiler production conditions characterized by high stocking density, severe occlusion, and constrained computational resources. In this context, the present study aimed to evaluate three lightweight object detection architectures for broiler monitoring and to determine their suitability for low-cost edge deployment in settings relevant to small and medium-sized producers. A novel dataset, publicly released through Zenodo to support reproducibility, was constructed from images acquired in both a prototype farm and a high-density commercial facility. These environments captured the visual complexity of intensive broiler production, where overlapping individuals and frequent occlusion challenge detection performance. YOLOv10s, Faster R-CNN, and EfficientDet-D0 were trained and evaluated for detection accuracy and computational efficiency. YOLOv10s achieved the best results, with a mean Average Precision (mAP) of 0.95, whereas Faster R-CNN and EfficientDet-D0 were less suitable for crowded scenes due to region proposal saturation and limited feature-extraction capacity. The selected model was further implemented on a Raspberry Pi 5, achieving a stable latency of 392.17 ms. These results demonstrate that YOLOv10s provides a robust balance between accuracy and efficiency for local broiler monitoring on affordable hardware, while also indicating that active thermal management is necessary to maintain operational stability under real-world conditions.