To address the security threats posed by unauthorized and erratic drone flights in low-altitude airspace defense and the technical challenge of real-time and precise early warning in complex environments, this paper proposes a real-time drone intrusion warning system that integrates the YOLOv10 object detection algorithm with the GLM4-V multimodal large model. The system hardware consists of high-sensitivity electro-optical detection equipment and a central computing unit, while the software integrates a client interaction platform and a multimodal data processing module. At the algorithmic level, YOLOv10 is employed for rapid drone localization , supplemented by the SORT algorithm for dynamic trajectory tracking. An innovative semantic description generation framework based on GLM4-V is designed, incorporating a structured prompt template to convert detection results (target location, information, and environmental background) into compressible text descriptions in real time. This significantly reduces the communication load in bandwidth-limited scenarios. Experimental results demonstrate that under extreme conditions—maximum detection range of 1 km and maximum flight speed of 23 m/s—the system achieves an identification accuracy of 97.6%. By forming a closed-loop process of "visual detection–semantic analysis–dynamic transmission," the proposed system effectively addresses the challenge of real-time early warning for drone intrusions in complex environments, offering a highly robust solution for low-altitude airspace security.