Developing a Low Resource Intensive Algorithm for Early Detection of Flying Swarm of Pest Birds

Abstract Granivorous pest birds, particularly the Red-billed Quelea, cause severe cereal crop losses in sub-Saharan Africa and disproportionately affect small-scale farmers. Traditional deterrence methods are increasingly ineffective, while modern deep-learning detectors often assume costly GPU resources. This paper proposes a low-resource hybrid detection pipeline that uses motion-based preprocessing to gate deep-learning inference. A Gaussian Mixture Model (GMM) and Motion History Images (MHI) continuously monitor motion and trigger YOLOv8-nano inference only when significant motion is detected. Experiments on a custom dataset of 1,380 swarm-annotated images achieved a test mAP@50 of 74.9%, precision of 77.1%, recall of 79.4%, and an average processing rate of 4.59 FPS on CPU-only hardware. CPU usage averaged 79.88% and dropped to 12.5% during static scenes, supporting the efficiency benefits of motion-gated inference for early-warning detection.