In the field of agriculture, the accurate detection of horses can not only improve the efficiency of ranch management, but also provide important data support for horse health monitoring and behavior analysis, and help promote the intelligent and sustainable development of agriculture. Aiming at the intelligent needs of modern agriculture and animal husbandry, this study proposes an improved horse detection model for YOLOv8. By introducing the GSConv module in the backbone network to strengthen the feature extraction capability, and using the VoV-GSCSP module to replace the C2f module in the neck network, the detection accuracy and computational efficiency are significantly improved. The experimental results show that combining the GSConv and VoV-GSCSP modules increases the precision of the model by 0.9%, the recall by 1.9%, and the mean average precision (mAP) by 1.9%, while the number of parameters and the computational effort are reduced by 5% and 11%, respectively. The improved model demonstrated higher accuracy and efficiency in the horse detection task and reduced false and missed detections, proving its potential application value in horse detection, which can provide a reference for intelligent horse management in the future.