The decline in cow conception rates in the Japanese cattle industry－from 68.7% in 1989 to 52.0% in 2022－poses a critical challenge to productivity. Major factors contributing to this decline include missed estrus detection due to labor shortages and the obscuration of estrus behaviors by environmental stressors such as heat stress. Conventional object detection methods are limited to capturing broad behavioral categories, such as mounting or physical contact, and do not account for subtle behavioral nuances, including mounting direction or vulva sniffing. To address these challenges, this study proposes a non-invasive method for fine-grained behavioral analysis based on fixed surveillance cameras and AI-driven pose estimation. By integrating YOLOv9 and DeepLabCut, we conducted detailed behavioral detection of mounting direction and specific contact points in Tosa Akaushi. The results indicated that all types of mounting behavior showed significant increases only on the day of estrus, suggesting that mounting behavior alone is insufficient for detecting potential signs of proestrus. In contrast, nose and forehead contact behaviors exhibited increasing trends on Days －3 and －1 relative to baseline levels, suggesting that these behaviors may serve as potential indicators of the proestrus phase. These findings suggest that detailed behavior quantification based on pose estimation may improve the identification of both proestrus and the day of estrus, thereby contributing to reproductive management for timely insemination.
