Abstract: Wearing personal protective equipment is an important means to ensure the safety of mine personnel. It is an important task of mine safety management to carry out mine personnel protective equipment monitoring. Coal mine underground environment is more complex, video surveillance is susceptible to noise, light and dust and other factors interference, resulting in the existing target detection methods for mine personnel protective equipment there are low detection accuracy, poor real-time, model complexity and so on, proposed an improvement of YOLOv8 real-time monitoring of mine personnel protective equipment method, known as DBE-YOLO. The DBE-YOLO model is first combined with deformable Convolution (DCNv2) in the CBS module of the benchmark model backbone network to form a DBS module. Making convolution deformable, when sampling, it can more closely detect the true shape and size of the object, more robust, It effectively improves its feature acquisition ability for targets of different scales. It is beneficial for the model to extract more feature information of personnel protective equipment and improve the model detection. Secondly, the weighted bidirectional feature pyramid mechanism (BiFPN) is integrated in the feature enhancement network. In the process of multi-scale feature fusion, the less efficient feature transmission nodes are deleted. Achieve a higher level of integration, the fusion efficiency of different scale features is improved. BiFPN also introduces a weight that can be learned. Helps the network learn the importance of different input features. Finally, WIoUv3 is used as the loss function of the model. By dynamically distributing gradient gain, Focus on ordinary anchor frame quality, In the process of model training, the harmful gradient generated by low quality anchor frame is reduced. The model performance is further improved. The experimental results show that DBE-YOLO model has a good effect in the monitoring of mine personnel protective equipment. The accuracy, recall and average accuracy were 93.1%, 93.0% and 95.8%, respectively. Compared with the benchmark model, it was increased by 0.8%, 2.9% and 2.9% respectively. Detection real-time improved to 65 f·s⁻¹, An increase of 8.3%, In addition, the number of parameters, floating point computation and model volume are 2 M, 6.6 G and 4.4 MB respectively. Compared with the original model, they were reduced by 33.3%, 18.5% and 30.2% respectively. The improved model is verified by using video surveillance of coal mine field operation. It effectively improves the problem of missing and false detection, It provides technical means for improving the operation safety of mine personnel.