在煤矿企业的生产过程中，为了避免不戴安全帽引发的安全事故，不仅要提高劳动者的安全防范意识，还应加大对佩戴安全帽行为的监督管理力度。目前，在矿井下的生产环境中，安全监测主要依赖于调度室监控人员观看视频监控数据来识别不安全行为。然而，这种方式耗时费力且误检率较高。更重要的是，无法实现全时段的持续监控，给生产带来了安全隐患。为此，本文开展了针对矿井下人员安全帽佩戴检测与跟踪算法的研究，主要研究内容如下：

针对矿井下所采集的图像中存在对比度低、光照不均，以及粉尘、水雾所导致的散射模糊等问题，提出一种基于自适应双通道先验的矿井下图像增强算法。采用融合暗通道与亮通道建立双通道先验模型，引入自适应权重系数来提高透射率图的精度，采用梯度导向滤波代替传统导向滤波对透射率图进行细化处理，并改进环境光值估算方法，提高算法增强效果。最后通过与Retinex算法、多尺度视网膜增强算法（Multi-Scale Retinex, MSR）以及暗通道先验等主流算法进行对比实验，结果表明本文提出的算法有效的提高了图像的信息量，增强了图像对比度，同时减轻井下光照不均以及粉尘造成的图像质量不高的问题。

针对煤矿井下复杂环境所导致的安全帽目标检测算法精确度低、漏检率高等问题，提出一种基于YOLOv5s改进的矿井人员安全帽佩戴检测算法。首先在原三层输出层的基础上，利用级联网络，添加一个专门针对于小目标检测的输出层；然后引入CBAM注意力机制，使网络能够更加精确的提取到目标特征信息；在其预测部分，改进损失函数提高网络检测精度；最后在主干网络中引入轻量型网络Shufflenetv2，降低网络的计算量和参数量。将改进后算法在CUMT-HelmeT数据集中进行对比实验，相较于SSD、Faster RCNN、YOLOv5s以及YOLOv7算法mAP@0.5分别提高了11.89%、10.62%、6.32%和1.39%，同时将模型参数量降低了7.6%。

本文针对现有DeepSORT目标跟踪算法在矿井环境中对于小目标安全帽的跟踪效果差，精度低等问题，通过提升检测器的准确度，从而改善跟踪算法的性能。为了使跟踪算法更好适应矿井下环境，采用扩展卡尔曼滤波对DeepSORT进行改进，提升了算法在矿井下安全帽跟踪任务中的适应性与鲁棒性，同时引入广义交并比GIoU优化匹配网络，提升模型目标跟踪准确度。实验结果显示，该算法能较好地解决跟踪过程小目标安全帽遮挡后丢失问题，并且能够满足井下工作人员安全帽跟踪的实时性要求。

In the production process of coal mining enterprises, in order to avoid safety accidents caused by not wearing safety helmets, it is not only necessary to improve the safety prevention awareness of workers, but also to increase the supervision and management of wearing safety helmets. At present, safety monitoring in underground production environments mainly relies on the dispatch room monitoring personnel watching video surveillance data to identify unsafe behaviors. However, this method is time-consuming, laborious, and has a high false detection rate. More importantly, the inability to achieve continuous monitoring throughout the entire time period has brought safety hazards to production. Therefore, this article conducted research on the detection and tracking algorithms for the wearing of safety helmets by underground personnel. The main research content is as follows:

A mining image enhancement algorithm based on adaptive dual channel priors is proposed to address the issues of low contrast, uneven lighting, and scattering blur caused by dust and water mist in the images collected from underground mines. We establish a dual channel prior model by integrating dark and bright channels, introduce adaptive weight coefficients to improve the accuracy of the transmittance map, use gradient guided filtering instead of traditional guided filtering to refine the transmittance map, and improve the method of estimating ambient light values to enhance the algorithm's enhancement effect. Finally, comparative experiments were conducted with mainstream algorithms such as Retinex algorithm, Multi Scale Retinex (MSR) algorithm, and dark channel prior. The results showed that the proposed algorithm effectively improved the information content of images, enhanced image contrast, and alleviated the problem of uneven underground lighting and low image quality caused by dust.

A safety helmet wearing detection algorithm for mine personnel based on YOLOv5s improvement is proposed to address the issues of target occlusion caused by complex underground environments in coal mines, as well as low accuracy and high miss rate of small target safety helmet detection algorithms. Firstly, on the basis of the original three output layers, a cascaded network is used to add an output layer specifically designed for small object detection; Then, the CBAM attention mechanism is introduced to extract key image features more accurately; In its prediction section, improve the loss function to enhance the network detection accuracy; Finally, a lightweight network Shufflenetv2 is introduced into the backbone network to reduce the computational and parameter complexity of the network. The improved algorithm outperforms SSD, Faster RCNN, YOLOv5s, and YOLOv7 algorithms in the CUMT Helmet dataset mAP@0.5 Improved by 11.89%, 10.62%, 6.32%, and 1.39% respectively, while reducing the number of model parameters by 7.6%.

This article addresses the issues of poor tracking performance and low accuracy of the existing DeepSORT target tracking algorithm for small target safety helmets in mining environments. By improving the accuracy of the detector, the performance of the tracking algorithm can be improved. In order to better adapt the tracking algorithm to the underground environment, an extended Kalman filter was used to improve DeepSORT, enhancing the adaptability and robustness of the algorithm in the underground safety helmet tracking task. At the same time, a generalized intersection union ratio (GIoU) was introduced to optimize the matching network and improve the accuracy of the model's target tracking. The experimental results show that the algorithm can effectively solve the problem of loss of safety helmets after small target occlusion during the tracking process, and can meet the real-time requirements of safety helmet tracking for underground workers.