414 Enhancing estrus detection in group-housed dairy cattle: Evaluating computer vision against traditional and sensor-based methods.

Abstract Recent advancements in computer vision have the potential to revolutionize animal management in the dairy industry. As a cost-effective alternative to activity monitors, computer vision reduces the number of devices required per animal. The present study compares computer vision techniques (CV) with traditional visual observation (TVO) and activity monitors (ACT) for estrus detection. Ten lactating Holstein Friesian cows housed in a single pen were examined for estrus over a ten-day trial period. Twelve hours prior to trial initiation, cows were fitted with an AfiTag II activity monitor (Afimilk) and synchronized 60 hours later via a single injection of Lutalyse (Dinoprost tromethamine 5mg/mL; Zoetis). Four corner mounted cameras captured continuous 1080p resolution footage, resulting in a total of 960 hours of video for post hoc data analysis. All video was manually reviewed for positive identification of behavioral estrus based on a pre-defined threshold behavioral score. TVO was conducted by trained observers over three 15-minute periods per day beginning at 03:15, 11:45, and 15:15, with expression of estrus behaviors quantified throughout each day. ACT data was manually recorded every 12 hours and included steps (S) and lying bouts (LB). Differences in activity were calculated each collection period, with a two-fold increase identified as a positive indication of behavioral estrus based. YOLOv10s was used as the CV base model. Sixty-two hours of estrus-related footage were selected and extrapolated into 223,200 frames. From these extrapolated frames, 1,152 were manually labeled for estrus behaviors. Eighty percent of the labeled frames were used for training, while twenty percent were used for validation. Rectal temperature (RT), and temperature-humidity index (THI) were recorded throughout the trial period to assess environmental effects. Statistical analyses were performed using the PROC MIXED procedure in SAS software. LB and S were not significantly influenced by THI (LB: P = 0.48, S: P = 0.67), RT (LB: P = 0.37, S: P = 0.51), or each other (LB: P = 0.97, S: P = 0.97); however, they varied significantly by cow (LB: P < 0.01, S: P = 0.03) and hour post-injection (LB: P < 0.01, S: P < 0.01). Precision, recall, and F1 scores were calculated for each estrus detection method. TVO had 100% precision, 25% recall and an F1 score of 40%. ACT provided 87.5% precision, 70% recall, and an F1 score of 77.8%. CV achieved 95.9% precision, 91.4% recall, and an F1 score of 93%. Though TVO had perfect precision, its low recall score reflects many false negatives. To this end, modern technologies appear superior to TVO, with CV dominant over ACT. Given its greater performance and reduced reliance on per-animal devices, computer vision presents a cost-efficient and accurate estrus detection solution for dairy producers.