针对煤矿带式输送机场景存在尘雾干扰严重、背景环境复杂、人员尺度多变且易遮挡等因素导致人员入侵危险区域识别准确率不高等问题，提出一种基于改进YOLOv8模型的井下人员入侵带式输送机危险区域智能识别系统。改进YOLOv8模型通过替换主干网络C2f模块为C2fER模块，加强模型的细节特征提取能力，提升模型对小目标人员的识别性能；通过在颈部网络引入特征强化加权双向特征金字塔网络（FE−BiFPN)结构，提高模型的特征融合能力，从而提升模型对多尺度人员目标的识别效果；通过引入分离增强注意力模块(SEAM)增强模型在复杂背景下对局部特征的关注度，提升模型对遮挡目标人员的识别能力；通过引入WIoU损失函数增强训练效果，提升模型识别准确率。消融实验结果表明：改进YOLOv8模型的准确率较基线模型YOLOv8s提升2.3%，mAP@0.5提升3.4%，识别速度为104帧/s。人员识别实验结果表明：与YOLOv10m，YOLOv8s−CA、YOLOv8s−SPDConv和YOLO8n模型相比，改进YOLOv8模型对小目标、多尺度目标、遮挡目标的识别效果均更佳，识别准确率为90.2%，mAP@0.5为87.2%。人员入侵危险区域实验结果表明：井下人员入侵带式输送机危险区域智能识别系统判别人员入侵危险区域的平均准确率为93.25%，满足识别需求。

To address challenges such as severe dust and fog interference, complex background environments, and variable personnel scales with frequent occlusions in coal mine belt conveyor scenarios, which resulted in low accuracy in recognizing personnel intrusions into hazardous areas, an intelligent recognition system based on an improved YOLOv8 model was proposed. The improved YOLOv8 model enhanced detailed feature extraction by replacing the C2f module in the backbone network with the C2fER module, which improved recognition performance for small targets. The Feature Enhancement Weighted Bi-Directional Feature Pyramid Network (FE-BiFPN) structure was introduced into the neck network to strengthen feature fusion capabilities, thereby enhancing recognition of multi-scale personnel targets. The Separated and Enhancement Attention Module (SEAM) was incorporated to improve the model's attention to local features in complex backgrounds, which boosted its ability to recognize occluded personnel targets. Furthermore, the WIoU loss function was applied to enhance training outcomes, improving recognition accuracy. Ablation experiment results showed that the improved YOLOv8 model achieved a 2.3% increase in accuracy and a 3.4% improvement in mAP@0.5 compared to the baseline YOLOv8s model, with a recognition speed of 104 frames per second. Personnel recognition experiments demonstrated that, compared to YOLOv10m, YOLOv8s-CA, YOLOv8s-SPDConv, and YOLOv8n models, the improved YOLOv8 model delivered superior recognition performance for small, multi-scale, and occluded targets, achieving a recognition accuracy of 90.2% and an mAP@0.5 of 87.2%. Personnel intrusion experiments revealed that the intelligent recognition system achieved an average accuracy of 93.25% in identifying personnel intrusions into belt conveyor hazardous areas, satisfying recognition requirements.