由于煤矿井下环境复杂、光照分布不均匀，造成获取的视频信息中存在亮度低、模糊、光照不均等问题，从而导致在井下人员检测与跟踪过程中算法易出现漏检问题。因此，为保障井下人员安全，研究如何提高检测与跟踪的准确性对于保障煤矿生产过程中的安全具有重要意义。本文主要研究如下：

（1）采用改进的Zero-DCE算法实现低照度图像增强。通过添加噪声损失函数抑制算法增强过程中凸显的噪声问题；采用自适应伽马校正对增强后部分图像出现的欠曝光或过曝光现象进行二次处理。仿真与实验结果表明，该算法可有效抑制噪声，提高了图像对比度。

（2）采用改进的YOLOv7算法实现井下人员检测。针对算力约束限制模型部署问题，通过Ghost轻量化模块优化YOLOv7主干网络结构，从而压缩模型；对于压缩后模型精度下降问题，优化BiFPN特征金字塔网络结构，实现丰富的语义信息提取；串联形式嵌入ACMix注意力机制，提高小目标检测精度。仿真与实验结果表明，与基准YOLOv7算法相比，改进的算法模型压缩了21.7%，AP提升了1%，召回率提升了1.6%，精确度提升了1.4%，结果表明算法具有较好检测效果。

（3）采用改进的DeepSORT算法实现井下人员跟踪。为了解决因遮挡而产生的目标丢失或ID变换问题，选用OSNet全尺度网络优化浅层残差网络，提高表观特征提取能力；优化IOU匹配方式，采用CIOU匹配方式判断检测框与边界回归之间的匹配程度。仿真与实验结果表明，改进的DeepSORT跟踪算法MOTA提高了1.6%，MOTP提高了0.5%，IDs下降了42，增强了模型的鲁棒性。

视频监控中的井下人员检测与跟踪算法发挥了深度学习技术从数据中获得更深层次特征的优势，保障了井下人员作业时的安全。该方法不仅适用于煤矿井下等低照度空间，也对其他复杂场景的检测与跟踪具有一定的借鉴作用。

Due to the complex underground environment and uneven light distribution in coal mines, there are problems such as low brightness, blurred and uneven lighting in the obtained video information, resulting in the algorithm being prone to missed detection problems in the process of underground personnel detection and tracking. Therefore, in order to ensure the safety of underground personnel, it is of great significance to study how to improve the accuracy of detection and tracking to ensure the safety of coal mine production process. The main research of this thesis is as follows:

The improved Zero-DCE algorithm is used to achieve low-light image enhancement. By adding a noise loss function to suppress the noise problems highlighted during the process, the algorithm is enhanced. Adaptive gamma correction is used to perform secondary treatment of the lack of exposure or overexposure of the enhanced images. The results of simulation and experiments show that the algorithm can effectively suppress noise and improve image contrast.

The improved YOLOv7 algorithm is used to realize the detection of underground personnel. Aiming at the problem of computing power constraint limiting model deployment, the YOLOv7 backbone network structure is optimized by Ghost lightweight module to compress the model. For the problem of model accuracy degradation after compression, the BiFPN feature pyramid network structure is optimized to achieve rich semantic information extraction. The tandem form embeds the ACMix attention mechanism to improve the accuracy of small target detection. The results of simulation and experiments show that compared with the benchmark YOLOv7 algorithm, the improved algorithm model has compressed 21.7%, AP increases by 1%, the recall rate has increased by 1.6%, and the accuracy has increased by 1.4%. The results show that the algorithm has a good detection effect Essence.

The improved DeepSORT algorithm is used to realize underground personnel tracking. In order to solve the problem of target loss or ID transformation caused by occlusion, the OSNet full-scale network is selected to optimize the shallow residual network and improve the apparent feature extraction ability. Optimize the IOU matching method, and judge the degree of matching between the detection box and the boundary regression by the CIOU matching method. The results of simulation and experiments show that the improved DeepSort tracking algorithm MOTA increased by 1.6%, MOTP increased by 0.5%, and IDs decreased by 42, which enhanced the robustness of the model.

The detection and tracking algorithms in the well-inferiority personnel in video surveillance have exerted the advantages of deep learning technology to obtain deeper characteristics from the data to ensure the safety of underground personnel during the operation. This method is not only suitable for low-illuminating space under the coal mine well, but also has a certain reference effect on the detection and tracking of other complex scenes.