Abstract: Existing personnel intrusion detection models have large numbers of parameters, pose challenges for lightweight deployment, and exhibit low intrusion recognition accuracy in scenarios involving partial contact or boundary intrusion. To address these problems, an early warning method for underground personnel intrusion into belt conveyor hazardous areas based on an improved YOLOv5-seg model was proposed. In the neck, the improved model replaced some standard convolution modules with lightweight GSConv modules and replaced the original C3 modules with lightweight C3Ghost feature fusion modules, reducing the number of parameters and computational load during feature fusion. In the head, the Proto branch was optimized by reducing the number of prototype feature channels and the dimensionality of mask coefficients to reduce the parameters and computation required for mask generation. Some standard convolutions were also replaced with GSConv to improve feature processing efficiency. A candidate buffer zone triggering mechanism was introduced, and a buffer zone was constructed by expanding the manually annotated actual hazardous zone outward. Personnel were preliminarily screened based on the spatial relationship between their detection bounding boxes and the expanded buffer zone. Instance segmentation was triggered only for targets entering this buffer zone to generate corresponding human body masks. Personnel intrusion was then determined from the overlap between these masks and the actual hazardous zone. Experimental results showed that the improved YOLOv5-seg model maintained detection accuracy while reducing the number of parameters and computational load by 22.75% and 45.12%, respectively, compared with YOLOv5-seg. Its frame rate reached 117 frames/s, achieving a favorable balance between accuracy and real-time performance. After the candidate buffer zone triggering mechanism was incorporated, precision, recall, and mAP@0.5 increased by 1.58%, 1.59%, and 1.19%, respectively, and the frame rate increased by 4.06%, while the number of parameters and computational load remained unchanged. These results indicated that the mechanism improved accuracy and real-time performance without increasing model complexity. The intrusion criterion based on overlap between human body masks and the actual hazardous zone maintained high detection accuracy in backlit scenes, scenes with complex interference, and multi-target scenes under low illumination. The average missed-alarm rate was only 11.3%, representing reductions of 78.64% and 74.61% compared with the conventional criteria based on the midpoint of the bounding box bottom edge and the entire bounding box bottom edge, respectively.