A cross-temporal individual-level framework for identifying, associating, and counting multiple mounting events for dairy cow in estrus monitoring

In vision-based estrus monitoring, mounting behavior serves as a key indicator. However, reliance on a single mounting event may cause false positives because non-estrous cows may also exhibit mounting behavior. To address this issue, a cross-temporal individual-level framework was proposed to identify, associate, and count multiple mounting events of the same cow over time. The framework combines an improved You Only Look Once Version 11-nano (YOLO11n) model with an improved baseline variant of the HiT lightweight hierarchical vision transformer tracking model (HiT-Base). The artificial intelligence contribution comprises improved detection and tracking. For detection, Normalized Wasserstein Distance loss was integrated with Complete Intersection over Union loss to improve small-target detection, and a lightweight Golden Cudgel Block (GCBlock) was designed to enhance feature representation under complex farm conditions. For tracking, a plug-and-play detail-enhanced convolution (DEConv) module was designed to strengthen boundary and fine-grained features, improving target discrimination and tracking stability. The engineering application contribution lies in reducing reliance on isolated observations and providing more reliable behavioral evidence through cross-temporal association and counting of multiple mounting events. Experimental results showed that the proposed detection model achieved 96.42% average precision and 94.74% balanced F-measure (F1-score) for cow in estrus, with an inference speed of 265.96 frames per second. The proposed tracking model achieved an area under the curve of 70.76%, a normalized precision of 88.73%, and an average overlap of 70.74%. These results demonstrate the effectiveness of the proposed framework for real-time estrus monitoring and its potential for precision dairy farming.