Automatic milking robots require accurate teat positions and consistent teat identities to guide cup attachment. However, weak texture, low contrast, rotated teat poses, and large inter-frame displacement in low-frame-rate Time-of-Flight (ToF) images make this task challenging, while embedded hardware imposes strict computational constraints. This study presents a cow teat detection and tracking framework comprising a lightweight YOLO-IWAL detector and a Siamese network-based tracking algorithm. Built on YOLO11n-OBB, YOLO-IWAL incorporates efficient feature extraction, detail-preserving downsampling, lightweight multi-scale feature fusion, and an oriented detection head. On the detection test set, YOLO-IWAL increases precision from 99.49% to 99.69%, reduces the number of parameters from 2.65M to 1.85M and computational complexity from 6.6 to 3.4 GFLOPs, and improves inference speed from 1039.2 to 1556.9 frames per second compared with the baseline. The Siamese network combines teat appearance features with 3D spatial information through Fourier positional encoding, adaptive geometric encoding, and gated fusion, and then applies the Hungarian algorithm for inter-frame identity association. On the tracking test set, it achieves an inter-frame identity association accuracy of 74.59% and an identity F1 score of 93.24%. Edge deployment on the RK3588 platform achieves 22.9 frames per second for YOLO-IWAL and a tracking processing speed of 166.39 frames per second for the INT8-quantized Siamese network. Overall, the proposed methods provide an efficient perception solution for cow teat detection and cross-frame identity association in automatic milking robots.
Zhang et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: