Researchers in China led by Zhihua Diao have built a deep learning system that recognizes what each Holstein dairy cow is doing and tracks every individual identity in real time on modest hardware, combining a slimmed-down object detector with an improved tracking algorithm and testing it on video from two dairy farms in Henan Province. Published in Artificial Intelligence in Agriculture, the results show the system detects eating, standing, and lying with high accuracy while running fast enough for continuous monitoring on edge devices. The motivation is practical: manual observation is inefficient, resource-intensive, and limited by human subjectivity, while wearable sensors face high maintenance costs and potential discomfort. Camera-based monitoring faces challenges including varying illumination, cows blocking one another, and frequent posture changes causing false detections, missed detections, and interrupted tracks; most existing research pursues accuracy through heavier networks, which the authors say overlooks computational load undermining real-time efficiency. The researchers installed DS-2CD3T46WD-I3 network cameras at a height of 2.8 meters with a 45-degree downward angle, spaced 5 meters apart, recording 1920 × 1080 video at 30 frames per second and capturing roughly 10 Holstein cows per area between June and December 2024, sampled across morning, afternoon, and night. They extracted one frame every 20 frames with ffmpeg and split images at an 8:1:1 ratio, producing a YOLO-format dataset of 8,000 images