传统的目标检测算法识别矿工排队异常行为时需人工提取特征，检测时间长、检测精度低；基于卷积神经网络的目标检测算法在检测速度和精度上有所提升，但在遮挡、昏暗和光照不均等场景下的检测效果难以保障。针对上述问题，提出了一种改进YOLOv5s（HPI−YOLOv5s）模型，并将其用于矿工排队检测。HPI−YOLOv5s模型在YOLOv5s模型的基础上对路径聚合网络（PANet）进行改进，通过删除单个输入边节点、增加双向交叉路径，构建了一种双向交叉特征金字塔网络（BCrFPN）进行多尺度特征融合。鉴于手动设置阈值的标签分配策略鲁棒性不高，在自适应训练样本选择（ATSS）动态设置阈值的基础上，提出动态标签分配策略（ATSS_PLUS），更合理地评估候选样本的质量，动态设定每个真实目标的阈值，具有更高的检测精度和鲁棒性。通过半平面交法计算人脸框与所划定排队区域的相交面积，并将相交面积和人脸框面积之比与设置的阈值比较以判断矿工是否有序排队。实验结果表明：HPI−YOLOv5s模型比YOLOv5s模型的准确率提高了1.9%，权重大小减少了32%，参数量减少了6.9%，检测速度提高了7.8%，且针对遮挡、昏暗、光照不均的矿井图像，能够更准确地识别矿工排队情况。

Traditional object detection algorithms require manual feature extraction when recognizing abnormal behavior of miners queuing, resulting in long detection time and low detection precision. The object detection algorithm based on convolutional neural networks has improved detection speed and precision. But its detection performance is difficult to guarantee in scenarios of obstruction, dimness, and uneven illumination. In order to solve the above problems, an improved YOLOv5s (HPI YOLOv5s) model is proposed. It is used for miner queue detection. The HPI-YOLOv5s model improves the path aggregation network (PANet) on the basis of the YOLOv5s model. By deleting a single input edge node and adding bidirectional crossing paths, a bidirectional cross feature pyramid network (BCrFPN) is constructed for multi-scale feature fusion. Considering the low robustness of label allocation strategies with manually set thresholds, a dynamic label allocation strategy (ATSS-PLUS) is proposed based on adaptive training sample selection (ATSS) to dynamically set thresholds. It can reasonably evaluate the quality of candidate samples and dynamically set thresholds for each real object, resulting in higher detection precision and robustness. The method calculates the intersection area between the face frame and the designated queue area using the half plane intersection method. The method compares the ratio of the intersection area to the face frame area with the set threshold to determine whether the miners are queuing in an orderly manner. The experimental results show that the HPI-YOLOv5s model has an accuracy improvement of 1.9%, a weight reduction of 32%, a parameter reduction of 6.9%, and a detection speed improvement of 7.8% compared to the YOLOv5s model. Moreover, it can more accurately recognize the queuing situation of miners in obstruction, dimness, and uneven illumination mine images.