Multimodal Detection Models for Poultry Fraud Monitoring on Jetson Nano

Abstract

This study defines an indoor commercial poultry-house scenario with no Global Positioning System (GPS) signal, variable bird density, illumination shifts, occlusion, and normal versus fraud episodes characterized as abnormal poultry population behavior (an unauthorized deviation between observed bird count and expected inventory baseline). We evaluate an unmanned aerial vehicle (UAV) to an edge-computing pipeline on Jetson Nano by comparing three models: You Only Look Once version 11 (YOLOv11) with red-green-blue (RGB) input, YOLOv11 with RGB and thermal late fusion, and a convolutional neural network (CNN) backbone with a support vector machine (SVM) classifier. The dataset contains 12,000 frames with synchronized RGB-thermal augmentation to preserve modality alignment. Evaluation covers mean Average Precision (mAP), precision, recall, F1-score, counting errors via mean absolute error (MAE) and root mean square error (RMSE), and edge metrics including frames per second (FPS), latency, and memory. YOLOv11 RGB+thermal records mAP@0.5 of 0.94 (Table 4a), MAE of 1.4, and RMSE of 2.0 (Table 4b), compared with YOLOv11 RGB at 0.91, 1.8, and 2.5 and CNN-SVM at 0.85, 2.6, and 3.4 (Table 4a-4b). For edge throughput, CNN-SVM reaches 28 FPS, while YOLOv11 RGB reaches 18 FPS and YOLOv11 RGB+thermal reaches 14 FPS (Table 8). As a scenario study, these metric-supported results indicate that YOLOv11 RGB+thermal is accuracy-first, CNN-SVM is speed-first, and YOLOv11 RGB is a balanced option for real-time poultry fraud monitoring.