Accurate and realtime livestock data are crucial to developing modern animal husbandry, ensuring effective supply of animal products, and promoting ecosystem balance and sustainable development of grasslands. The acquirement of livestock data mainly relies on field surveys and grassroots’ reports. These data are laborious and non-real-time. In this study, a real-time monitoring system is developed based on browser/server architecture (
http://218.202.104.82:5806/vid). A deep-learning-based livestock detection model and a weight estimation model are developed. The system could detect and count livestock, and estimate their weight using unmanned aircraft system (UAS) live video streams. The livestock detection model is trained using 13803 UAS image tiles and video picture frames. The true positive rate, false positive rate, and loss positive rate of the model for cattle detection are 90.51%, 11.64%, and 9.49%, respectively. The true positive rate, false positive rate, and loss positive rate of the model for sheep detection are 91.47%, 7.04%, and 8.53%, respectively. The weight estimation model is built based on the head-body length and weight data collected in Inner Mongolia Autonomous Region and Qinghai Province, with accuracy of 90.28% and 90.00% for cattle and sheep weight estimation, respectively. The system utilizes UASs and deep learning technologies for livestock monitoring, having an expected application prospect in the fields of grassland supervision (including grazing prohibition and rest grazing), and assisting herdsmen in remotely monitoring their livestock.