随着煤矿少人无人化、智能化开采不断发展，智能监控视频作为数据源成为研究煤矿安全的重要途径。然而，由于煤矿粉尘与光照的干扰、人员对关节的自遮挡、人员之间的相互遮挡、人员在图像中作为小目标占比较少等问题，导致信息识别精度低、难度大。因此，开展煤矿井下综采工作面人员视觉检测和人员行为异常识别研究具有重要意义。本文通过对图像进行去雾、抑制光照的方法改善图像的质量作为扩充数据集，同时改进网络结构并训练人体关键点检测模型、设置姿态优先级动作识别网络有效分析当前矿工的主要姿态。主要研究内容如下：

针对煤矿综采工作面高光、大雾使图像特征曝光、减弱等问题，提出了基于高光抑制的图像处理方法，引入亮度调节因子对彩色图像三个通道亮度分别进行抑制；将调节后的图像转化到HSV色彩空间内去除入射光、保留反射光并进行亮度均衡；采用基于改进引导滤波的暗通道先验算法进行去雾，引入伽马校正函数均衡图像亮度值。将本文算法与Retinex光照处理算法同时处理图像，并进行对比分析，采用信息熵、均值、标准差以及空间频率对处理后图像的信息、亮度、边缘信息及清晰度进行质量评估。实验结果表明，经过本文算法处理后的图像，不仅保留了原有图像特征，同时抑制了图像亮度、提高了图像清晰度及边缘信息。将此方法处理后的图像与原图同时作为数据集，既实现了数据的扩充，也有利于模型学习有用特征。

针对煤矿井下综采工作面人员在图像中占比小、恶劣环境对图像的影响较大以及现有的模型对此类环境下图像的学习精度较差等问题，提出基于yolov7x融合CBAM注意力机制的检测模型。该模型通过重参数化、高聚合以及辅助训练等模块来加速训练过程中的收敛速度，同时通过融合注意力机制来提升模型的检测精度。通过对人员及其关键点进行标注自制数据集，同时将经过去雾以及高照度光照抑制的图片也作为标注像，最后将共计一万张图像的自制数据集使用改进后的模型进行训练。实验结果表明，训练后的模型mAP值为97.5%、精确度值为94.5%、召回率的值为92.6，且模型在训练到第68后收敛。

针对在采煤过程中人员姿态多变、人员不规范操作等问题，提出基于改进yolov7x训练模型的人员姿态优先级的行为异常检测算法。在跟踪人员的过程中，通过训练模型检测出图像中的目标，并将目标的关键点同时解算出来，同时根据关键点解算的结果按照关节之间的相互关系分为多个局部关节小组，解算每个关节组的姿态，并将所有关节组的姿态按照姿态的重要程度设置优先级，即对人体行为影响最大或者造成结果最严重的姿态设置高优先级输出。最后将输出的结果与已知人员行为做比较，以便用来验证人体行为的正确性。

搭建人员的视觉跟踪及姿态异常检测嵌入式实验平台，对采集的煤矿井下综采工作面视频图像的人员进行实时性跟踪，并提取人员关键点信息判断人员是否处于异常状态，并根据先验信息对异常姿态进行验证。实验结果表明，本文提出人员的视觉跟踪及异常检测系统初步满足实时性，准确率高，基本可以实现人员的巡检要求，对综采面的自动化检测具有重要意义。

As coal mines increasingly move towards automation and intelligence, intelligent surveillance videos have become a crucial data source for studying mine safety. However, challenges such as interference from coal dust and lighting, self-occlusion of joints by personnel, mutual occlusion between personnel, and the relatively small size of personnel in images have led to low accuracy and high difficulty in information recognition. Therefore, conducting research on visual detection of personnel in underground coal mining faces and identifying abnormal behaviors is of significant importance.

This paper addresses these challenges by improving image quality through defogging and light suppression methods to expand the dataset. Additionally, it enhances network structures, trains human keypoint detection models, and establishes posture-priority action recognition networks to effectively analyze the primary postures of miners. The main research contents are as follows.

To address the issues of overexposure and feature attenuation caused by high light and heavy fog at fully mechanized coal mining faces, a high light suppression-based image processing method is proposed. This method introduces a brightness adjustment factor to separately suppress the brightness of the three channels of the color image. The adjusted image is then converted to the HSV color space to remove incident light, retain reflected light, and balance brightness. An improved guided filter-based dark channel prior algorithm is used for defogging, and a gamma correction function is introduced to balance the image brightness values. The proposed algorithm is compared with the Retinex illumination processing algorithm, and the images are analyzed using information entropy, mean, standard deviation, and spatial frequency to evaluate the quality in terms of information, brightness, edge information, and clarity. Experimental results show that the images processed by this algorithm not only retain the original image features but also reduce image brightness, and improve image clarity and edge information. Using these processed images along with the original images as the dataset not only achieves data augmentation but also facilitates the model's learning of useful features.

Detection Model Based on yolov7x and CBAM Attention Mechanism: Given the issues of the small proportion of personnel in images, significant environmental impact on images, and poor learning accuracy of existing models under such conditions, a detection model based on YOLOv7x integrated with the CBAM attention mechanism is proposed. This model accelerates the convergence speed during the training process through reparameterization, high aggregation, and auxiliary training modules, while the attention mechanism enhances the model's detection accuracy. A self-made dataset is created by annotating personnel and their key points, including images processed through defogging and high illumination suppression. The improved model is trained using a total of ten thousand annotated images. Experimental results show that the trained model achieves a mAP value of 97.5%, a precision of 94.5%, and a recall rate of 92.6%, with the model converging after 68 epochs of training.

To address issues such as the variability of personnel postures and non-standard operations during coal mining, a behavior anomaly detection algorithm based on an improved yolov7x training model is proposed. During the process of personnel tracking, the algorithm detects targets in the image using a trained model and simultaneously calculates the key points of these targets. Based on the results of key point calculations, the algorithm divides them into multiple local joint groups according to the interrelation of joints. It then calculates the posture of each joint group and prioritizes all joint group postures based on their importance. This prioritization ensures that postures with the most significant impact on human behavior or with the potential for the most severe consequences are assigned high priority outputs. Finally, the output results are compared with known human behaviors to validate the correctness of human actions.

Embedded Experimental Platform for Visual Tracking and Abnormal Posture Detection: An embedded experimental platform for personnel visual tracking and posture abnormality detection is built, which tracks personnel in real-time using video images collected from the underground fully mechanized mining face, extracts key point information of personnel, determines if personnel are in an abnormal state, and verifies abnormal postures based on prior information. Experimental results show that the proposed visual tracking and abnormal detection system for personnel preliminarily meets real-time requirements with high accuracy, essentially fulfilling personnel inspection requirements and holding significant importance for the automated detection of fully mechanized mining faces.