Edge Computing in Commercial Poultry: Minimum Artificial Intelligence Inference Requirements for Real-Time Broiler Harvest Count Verification

Abstract

Commercial poultry harvesting in many Philippine farm settings still relies on manual bird handling, weighing, and count verification procedures. Although computer vision can support these tasks, practical use depends not only on detector capability but also on whether inference can run reliably on affordable edge hardware near the point of operation, especially where cloud connectivity is limited or unstable. This study evaluates the minimum practical hardware conditions needed to support local broiler harvest count verification using a trained YOLO-based oriented bounding box workload.

The detector is treated as a fixed deployment workload, while the study focuses on hardware feasibility under realistic farm-side conditions. A live RTSP harvest-camera stream is processed locally while latency, achieved frames per second, memory use, processor utilization, and thermal behavior are recorded. Benchmarks on four CPU-only platforms show a wide separation in practical suitability. Effective throughput ranged from 4.44 to 24.73 frames per second in short runs and from 4.58 to 17.29 frames per second in sustained 10-minute runs. The strongest desktop platform maintained the best throughput with clear thermal margin, the laptop and mid-tier desktop remained usable but thermally strained, and the older 4-core desktop approached a practical lower bound. These results show that deployment suitability depends on sustained throughput and thermal stability rather than raw inference speed alone. The study therefore contributes a deployment-centered view of edge- computing sufficiency for broiler harvest verification.