Lightweight Visual Detection and Dynamic Tracking for Pigeon Egg Inspection in Caged Pigeon Farming.

Manual inspection in large-scale pigeon farms is inefficient and often misses critical targets. In addition, recognition results are difficult to link to physical cage locations in real time. Here, we develop an intelligent inspection and localization system that integrates an improved lightweight YOLO model with QR-code-based tracking. QR codes are deployed along the inspection route as spatial anchors. Base detection models are combined with the ByteTrack algorithm to establish a dynamic mapping among video frames, cage numbers and detected targets. To improve the detection of small pigeon eggs caused by interference from metal cage meshes, we further design a lightweight YOLO-PEDI (Pigeon Egg Detection Inspection) model. Ghost modules replace standard convolutions to reduce computational cost. CBAM is introduced to enhance feature extraction in complex backgrounds. The newly designed model enables simultaneous identification of egg number and egg condition, including normal and broken eggs. The proposed method achieves an mAP50 of 98.1%, with only 1.53 million parameters and an inference time of 0.8 ms. Field tests show a cumulative egg-counting accuracy of 80.0% and a broken egg detection rate of 98.0%. These results demonstrate the potential of the proposed system for intelligent inspection in pigeon farming and provide a practical route towards precise traceability and digital production management.