随着计算机技术和煤炭工业生产力的不断提高，煤矿开采朝着智能化、无人化方向发展，对矿井进行三维建模是评价煤矿开采智能化程度的重要参考指标。基于视觉的三维重建技术是常用的三维建模方法之一，因其成本低、操作便捷的优势逐渐成为当前的研究热点。在基于视觉的三维重建技术中，运动恢复结构（Structure from Motion, SFM）因能自动地完成相机追踪与运动匹配的特点得到了广泛的应用，但在真实场景下，SFM三维重建方法易受到图像质量、匹配算法等多种因素的影响。尤其是针对矿井环境而言，井下光照不均、低照度及粉尘等问题使得三维重建难度大大增加。因此，本文从提高图像质量、优化匹配算法两方面对矿井环境下的三维重建方法进行了研究。具体工作如下：

（1）针对矿井复杂环境下，采集的图像易受光照不均、低照度、粉尘等因素而导致图像存在失真、对比度低的问题，提出一种井下图像增强算法。该算法首先通过改进同态滤波算法提高图像清晰度，其次采用改进直方图均衡化算法提高图像对比度，最后与传统同态滤波算法相比，实验结果表明，本文图像增强算法所获取的图像信息熵相较于高斯型同态滤波提高了0.25比特，能够为后续图像匹配提供更加丰富的特征纹理信息。

（2）针对现有特征提取与匹配算法在矿井环境下特征点提取不准确、匹配精度低的问题，提出了基于图论的改进ORB算法和基于图像增强的改进AKAZE算法。通过构建图的方式和图像增强的方法对传统的ORB算法和AKAZE算法进行了不同程度的改进，实验结果表明，前者能够获取质量更高的特征点，且数量比传统的ORB算法至少增加了1.4倍；而后者能够有效的提高匹配正确率，相较于SIFT、AKAZE算法的平均正确率分别提高了18.03％、8.76%，能够为后续的三维重建提供更加精确的匹配点对。

（3）针对传统SFM算法中，使用SIFT算法进行特征提取与匹配会丢失大量边界信息的问题，提出了基于改进SFM算法的三维重建方法，并实现了对矿井环境中目标物体的三维建模，最后与传统基于SIFT算法的三维重建方法相比，本文所提出的改进SFM算法在保证效率的同时，其点云数量提高了1.84倍，RMSE减小了37.4%

With the continuous improvement of computer technology and the productivity of the coal industry, coal mining is developing towards an intelligent and unmanned direction. 3D modeling of mines is an important reference index for evaluating the intelligence of coal mining. Vision-based 3D reconstruction technology is one of the commonly used 3D modeling methods, and has gradually become a current research hotspot due to its low cost and convenient operation. In vision-based 3D reconstruction technology, Structure from Motion (SFM) has been widely used because it can automatically complete camera tracking and motion matching. Quality, matching algorithm and other factors. Especially for the mine environment, the problems of uneven lighting, low illumination and dust in the underground make the 3D reconstruction much more difficult. Therefore, this paper studies the 3D reconstruction method in the mine environment from the aspects of improving the image quality and optimizing the matching algorithm. The specific work is as follows:

(1) Aiming at the problems of image distortion and low contrast caused by the uneven illumination, low illumination, dust and other factors of the collected images in the complex environment of the mine, an underground image enhancement algorithm is proposed. The algorithm firstly improves the image clarity by improving the homomorphic filtering algorithm, and then uses the improved histogram equalization algorithm to improve the image contrast. Finally, compared with the traditional homomorphic filtering algorithm, the experimental results show that the image information entropy obtained by the image enhancement algorithm in this paper is improved. Compared with Gaussian homomorphic filtering, it is increased by 0.25 bits, which can provide richer feature texture information for subsequent image matching.

(2) Aiming at the problems of inaccurate feature point extraction and low matching accuracy of existing feature extraction and matching algorithms in mine environment, an improved ORB algorithm based on graph theory and an improved AKAZE algorithm based on image enhancement are proposed. The traditional ORB algorithm and the AKAZE algorithm have been improved to varying degrees by the method of constructing the graph and the method of image enhancement. The experimental results show that the former can obtain higher quality feature points, and the number is at least 1.4 more than that of the traditional ORB algorithm. The latter can effectively improve the matching accuracy rate, which is 18.03% and 8.76% higher than the average accuracy of SIFT and AKAZE algorithms, respectively, which can provide more accurate matching point pairs for subsequent 3D reconstruction.

 (3) In the traditional SFM algorithm, using the SIFT algorithm for feature extraction and matching will lose a lot of boundary information, a 3D reconstruction method based on the improved SFM algorithm is proposed, and the 3D modeling of the target object in the mine environment is realized. Finally, compared with the traditional 3D reconstruction method based on the SIFT algorithm, the improved SFM algorithm proposed in this paper not only ensures the efficiency, but also increases the number of point clouds by 1.84 times and reduces the RMSE by 37.4%.