煤矿井下环境复杂，监控设备在成像时受到诸多因素的干扰，导致呈现的图像模糊，因此使用监控视频检测井下矿工时会存在误检率较高等问题。为保障矿井人员安全，基于矿井监控视频的井下人员检测对于煤矿安全生产具有重要意义。本文主要的工作如下：

（1）针对煤矿井下监控设备得到的图像较为模糊且照度分布不均匀等问题，提出 一种基于同态滤波的矿井监控视频图像增强算法。首先对原有的同态滤波做出改进，在引入单参数同态高通滤波的基础上，结合改进的同态低通滤波进行加权融合。然后将 RGB 空间下的图像转换至 HSV 空间中，保持 H（色调通道）不变，对 V（亮度通道） 采用限制对比度自适应直方图均衡化算法进行亮度校正，根据亮度变化对 S（饱和度通 道）进行自适应非线性增强。最后将图像反转回 RGB 空间，完成对矿井图像的增强。 实验结果表明，本文算法增强了图像对比度，且能去除图像中的大部分噪声，使得图像 具有良好的保真性。

（2）针对煤矿井下多尺度、小目标矿工检测精度较低的问题，以 YOLOv5 模型为基础，提出YOLOv5-E 矿井人员检测算法。首先采用 EfficientNetV2 结合 CA 注意力机 制作为特征提取网络，以便更好的提取井下人员特征。其次使用 Bi-FPN 网络作为特征 融合网络，将不同尺度之间的特征充分利用，最终使得矿井人员检测算法在满足实时检测的条件下还具有较高的精度。实验结果表明，本文检测算法在自制井下行人数据集中得到的精度值为 90.6%，每秒帧率达到 72FPS。

（3）为了验证本文预处理算法的有效性，使用本文检测算法分别检测增强前后矿 井监控视频图像。实验验证增强后图像可使得检测精度提高 1%。为了验证本文检测算 法的检测性能，将本文矿井人员目标检测算法与当前主流目标检测算法进行对比实验。 实验结果表明，本文算法对矿井人员的检测精度高，实时性好。为了进一步检验本文检 测算法的泛化能力，使用公共数据集检验本文算法对人体目标的检测效果。实验结果表 明，本文算法具有良好的泛化性。

本文方法对矿井监控视频图像增强和井下人员检测有一定参考价值，为煤矿的安全 生产提供了有益支撑。

Underground coal mines are a complex environment, and surveillance equipment is subject to interference from many factors when imaging, resulting in blurred images, so there is a high rate of false detection when using surveillance video to detect underground miners. In order to ensure the safety of mine personnel, underground personnel detection based on mine surveillance video is of great importance to the safe production of coal mines. The main work of this thesis is as follows:

(1) A homomorphic filtering-based image enhancement algorithm is proposed for the blurred images and uneven illumination distribution obtained from underground monitoring equipment in coal mines. Firstly, the original homomorphic filtering is improved by introducing a single-parameter homomorphic high-pass filter, combined with an improved homomorphic low-pass filter for weighted fusion. The image in RGB space is then converted to HSV space, keeping the H (Hue Channel) unchanged, using a restricted contrast adaptive histogram equalisation algorithm for V (Value Channel) for luminance correction, and adaptive non-linear enhancement for S (Saturation Channel) according to luminance changes. Finally, the image is inverted back to RGB space to complete the enhancement of the mine image. The experimental results show that the algorithm in this thesis enhances the image contrast and can remove most of the noise in the image, which makes the image have good fidelity.

(2) To address the problem of low accuracy in detecting multi-scale and small target miners in underground coal mines, the YOLOv5-E mine personnel detection algorithm is proposed based on the YOLOv5 model. Firstly, EfficientNetV2 combined with CA attention mechanism is used as a feature extraction network for better extraction of underground personnel features. Secondly, the Bi-FPN network is used as a feature fusion network to make full use of the features between different scales, which finally makes the mine personnel detection algorithm have high accuracy while meeting the condition of real-time detection. The experimental results show that the accuracy value obtained by the detection algorithm in this thesis is 90.6% in a homemade underground pedestrian dataset, with a frame rate per second of 72 FPS.

(3) In order to verify the effectiveness of the pre-processing algorithm in this thesis, the mine monitoring video images before and after enhancement were detected using the detection algorithm in this thesis. The enhanced images were experimentally verified to improve the detection accuracy by 1%. In order to verify the detection performance of the detection algorithm in this thesis, the mine personnel target detection algorithm in this thesis is compared with the current mainstream target detection algorithm. The experimental results show that the algorithm in this thesis has high detection accuracy and good real-time performance for mine personnel detection. In order to further test the generalization ability of the detection algorithm in this thesis, the detection effect of this thesis's algorithm on human targets is examined using a public data set. The experimental results show that the algorithm in this thesis has good generalisation.

The method in this thesis has some reference value for mine video image enhancement and underground personnel detection, and provides useful support for the safety production of coal mines.