Abstract Traditional visual estrus detection can be labor-intensive and impractical for continuous monitoring under commercial conditions. Although wearable biosensors provide automated alternatives, their high-cost limits widespread adoption. Advancements in computer vision present a promising, cost-effective approach to this issue. This study developed and evaluated a two-layer computer vision framework, integrating behavioral detection, individual animal identification, and multi-animal tracking to facilitate accurate and efficient estrus detection in dairy cows. Thirty Holstein-Friesian cows were monitored across three consecutive 10-day trial periods (n = 10 cows/trial) in a single freestall pen. Forty-eight hours after trial initiation, cows were synchronized using a single intramuscular injection of Lutalyse (Dinoprost tromethamine, 5 mg/mL; Zoetis). Continuous video was captured using four strategically positioned 1080p cameras, yielding 2,880 total hours of behavioral data. From this dataset, 95 hours containing observable estrus-related behaviors were selected and converted into individual image frames. A total of 135,850 frames were extracted, while 3,747 frames were used to train the estrus detection (ED) model and 1,229 frames were used to train the identification (ID) model. Each dataset was partitioned into training and validation subsets using an 80:20 ratio. Two YOLOv10s models were trained for 100 epochs with a batch size of 16 and an input resolution of 640 × 640 pixels. The first model (ED) detected estrus-related behaviors, specifically chin-resting and mounting. When a positive estrus event was identified, a second model (ID) was triggered to determine the individual identity of cows within the detected region of interest. Detected individuals were then tracked across frames using the DeepSORT algorithm to maintain consistent identities, and consecutive detections of the same event were merged into single occurrences. The first, midpoint, and final frames of each event were retained for verification. Model performance was evaluated using a 60-minute video containing both estrus and non-estrus behaviors to assess accuracy and robustness across activity contexts. Model performance was evaluated against manually annotated ground-truth data. The ED layer correctly identified 85 true positives, with 71 false positives and 23 false negatives, resulting in a precision of 54%, recall of 79%, and an F1-score of 64%. The ID layer achieved 86 true positives, 70 false positives, and 23 false negatives, corresponding to a precision of 55%, recall of 79%, and an F1-score of 65%. Overall, the two-layer YOLOv10s–DeepSORT framework demonstrated reliable performance in detecting and tracking estrus behaviors in group-housed dairy cows. These results indicate that the combined approach is capable of recognizing estrus events and maintaining cow identity across frames under commercial conditions. While the models occasionally produced false positives, their strong recall suggests robustness for continuous behavioral monitoring, establishing a promising foundation for scalable, automated, cow-specific estrus detection systems in precision dairy management.
Clayton et al. (Wed,) studied this question.