Аңдатпа
Traditional linear evaluation of cattle conformation is widely used in breeding and herd management but is labor-intensive and dependent on operator measurements. This pilot study presents and evaluates a hierarchical multi-camera 3D reconstruction framework based on an Edge AI architecture using OAK/RVC4 devices for non-contact cattle phenotyping. The system integrates synchronized RGB-D acquisition, YOLO11-based detection and segmentation, depth reconstruction, multi-view point-cloud registration, weighted fusion, and RFID-based individual identification. The pilot validation showed mean absolute percentage errors (MAPE) of 0.72%, 0.82%, and 0.70% for withers height, rump height, and back height, respectively, based on four breeding bulls from a breeding center. The measured processing cycle was 535 ms, including synchronized capture, YOLO11 inference, depth generation, and weighted ICP-based fusion. Ischial width showed greater variability than the other evaluated traits, with a maximum relative error of 4.8% for one breeding bull. The tested configuration included active IR projection to support depth acquisition on low-texture surfaces; however, the present pilot design does not provide a controlled ablation experiment to quantify the independent contribution of IR projection. The results demonstrate the technical feasibility of the integrated reconstruction and morphometric workflow for breeding bulls, while larger studies with repeated scans and independent validation cohorts are required to assess repeatability, agreement, generalizability across breeds, and performance under varying environmental conditions.