<p>煤炭是我国最主要的一次能源。随着透明工作面建设的持续推进，激光雷达在综采工作面发挥着越来越重要的作用。点云配准是综采工作面三维激光扫描建模的关键技术。综采工作面的尘雾环境影响激光雷达测量精度和测量范围，导致点云配准存在精度低和误匹配率高的问题。通过单目相机与激光雷达的数据融合，能避免单一传感器使用的局限性，提高点云配准的精度和鲁棒性。针对综采工作面环境特点，对单目相机与激光雷达数据融合的点云配准方法进行深入研究，主要包括以下内容：</p> <p>针对综采工作面单目相机与激光雷达在线标定存在的标定板提取效率低的问题，研制单目相机与激光雷达数据融合系统，设计井下在线标定板。在此基础上提出基于改进区域生长的标定板点云提取算法，首先采用体素随机法对点云进行降采样；其次采用欧氏距离和角度距离结合的方法实现点云的快速聚类；然后采用最小二乘法求解法向量计算平面参数，生成初始种子；最后对原点云采用区域生长方法进行标定板点云的精细化提取。标定板点云提取平均时间为32ms，提取速度比RANSAC算法提高12.1%，重投影误差小于4像素，旋转误差小于0.7&deg;，平移误差小于6mm。结果表明在保证精度的同时提高标定板点云检测效率。</p> <p>针对综采工作面图像增强存在的亮区域增强过度、暗区域细节增强不足的问题，提出双通道块匹配的图像增强算法。首先对原始图像数据进行暗通道和亮通道分析，估计环境光值和透射率；其次根据图像的非局部相似性对像素进行块匹配分组；然后通过小波变换将三维块转换成小波域系数，对系数进行硬阈值滤波；最后经过小波逆变换恢复增强后的图像。处理时间稳定在22ms以内，与传统图像增强算法相比较，将综采工作面图像的峰值信噪比提高2.85%，结构相似性提高1.31%，自然图像质量提高7.91%。结果表明提高综采工作面图像的清晰度和对比度，有效改善图像的质量。</p> <p>针对综采工作面激光雷达点云滤波存在的近场噪声过多和远场有效数据丢失问题，构建尘雾对激光雷达的影响模型。在此基础上提出基于扇环邻域的综采工作面点云两次滤波算法。首先采用主成分分析法将三维点云数据进行降维处理，得到多组二维数据；其次对每组数据进行拉普拉斯运算，识别数据中的高频噪声，对数据进行第一次滤波；然后采用扇环邻域的统计分析，删除数据中的异常点，对数据进行第二次滤波；最后，将经过两次滤波后的数据恢复成三维点云。获得97.8%的滤波准确率，计算时间稳定在0.5s以内。对比统计滤波、半径滤波和动态半径滤波算法，结果表明在提高点云滤波的准确率和改善点云的结构化特征的同时，减少运行时间，有效抑制点云中的尘雾噪声。</p> <p>针对综采工作面激光雷达点云聚类存在的阈值依赖和密度不均匀问题，提出单目相机与激光雷达数据融合的基于参考平面投影的综采工作面点云聚类算法，首先对增强后的图像采用梯度分位数双阈值进行边缘检测，得到二值边缘图像；其次对滤波后的点云采用随机采样一致性方法提取点云中的煤壁区域；然后根据单目相机和激光雷达的标定矩阵，将边缘图像投影到点云数据中，精细化提取煤壁边缘，得到图像边缘约束的煤壁点云；最后采用参考平面投影变换法对非煤壁点云进行聚类，得到点云簇类。煤壁分割准确率约为93.5%，平均聚类时间为58ms，对比Kmeans和A-DBSCAN算法，聚类结果DB指数平均降低7.25%，CH分数平均提升9.06%。结果表明能有效减少点云分割中的过分割和欠分割现象，提高聚类点云簇的类间差异性。</p> <p>针对综采工作面激光雷达点云配准存在的精度低误匹配率高的问题，提出基于点云簇的光强灰度矫正点云配准算法。首先计算点云簇的特征值，通过特征匹配得到第一组候选匹配点对；其次提取关键点，采用扩展特征匹配得到第二组候选匹配点对；然后根据激光的反射光强和灰度矫正候选匹配点对，将候选匹配点对分类处理，剔除距离大的匹配点对，保留距离小的匹配点对，调整距离中间的匹配点对；最后采用鱼群搜索算法求解变换矩阵。计算时间稳定在0.35s以内，对比GP-ICP算法和CPD算法，RMSE平均降低11.4%。结果表明有效降低点云的误匹配率，提高综采工作面点云配准的精度。</p> <p>采用研制的单目相机与激光雷达数据融合系统，结合提出的算法，在西安科技大学煤炭主体专业综合实训中心的综采工作面进行现场实验。通过综采工作面机头点云、中间点云和机尾点云拼接实验，对本文所研究的理论和方法进行验证，取得较好的效果。本文提出的点云配准方法为透明工作面提供技术支撑，也为实现煤矿综采工作面无人化奠定理论基础。</p>

<p>Coal is the most important primary energy in China. With the continuous promotion of transparent working face construction, LiDAR is playing an increasingly important role in fully mechanized mining faces. Point cloud registration is a key technology for 3D modeling of fully mechanized mining faces. The dust and mist environment of the fully mechanized mining face affects the accuracy and range of LiDAR, resulting in low accuracy and high mismatch rate in point cloud registration. By integrating data from monocular cameras and LiDAR, the limitations of using a single sensor can be overcome, and the accuracy and robustness of point cloud registration can be improved. A point cloud registration method for data fusion between monocular cameras and LiDAR is studied based on the environmental characteristics of fully mechanized mining faces. The main content is as follows:</p> <p>In response to the problem of low extraction efficiency of calibration plates in online calibration of cameras and LiDAR, a data fusion system for monocular cameras and LiDAR was developed, and a underground coal mine calibration plate was designed. A calibration plate extraction algorithm based on improved region growth is proposed. Firstly, the voxel random method is used to down sampling the point cloud; Secondly, a combination of Euclidean distance and angular distance is used to achieve rapid clustering of point clouds; Then, the LSM is used to solve the normal vector to calculate the plane parameters and generate the initial seeds; Finally, the origin cloud is refined and extracted using the region growth method. The average extraction time is 32ms, and the extraction speed is 12.1% faster than the RANSAC algorithm. The reprojection error is less than 4 pixels, the rotation error is less than 0.7&deg;, and the translation error is less than 6mm. The results show that the detection efficiency is improved while ensuring accuracy.</p> <p>A dual channel block matching algorithm for image enhancement of fully mechanized mining faces is proposed to address the problems of excessive enhancement of bright areas and insufficient enhancement of dark areas in image enhancement. Firstly, perform dark and bright channel analysis on the original image data to estimate the ambient light value and transmittance; Secondly, block matching and grouping of pixels are performed based on the principle of non local similarity in the image; Then, the 3D blocks are transformed into wavelet domain coefficients through wavelet transform, and the coefficients are subjected to hard threshold filtering; Finally, the enhanced image is restored through inverse wavelet transform. The processing time is stable within 22ms. This algorithm improves the PSNR of image by 2.85%, SSIM by 1.31%, and natural image quality by 7.91%. The results show that improving the clarity and contrast of the images, effectively filters out dust and fog noise in the images, and improves the quality of the images.</p> <p>In response to the problems of excessive near-field noise and loss of far-field effective data in the LiDAR point cloud filtering of fully mechanized mining faces, a model of the impact of dust and mist on LiDAR has been established. On the basis of this, a point cloud twice filtering algorithm based on fan ring neighborhood is proposed. Firstly, principal component analysis is used to reduce the dimensionality of 3D point cloud data and obtain multiple sets of 2D data; Secondly, perform Laplace operations on each group of data to identify high-frequency noise in the data and perform the first filtering on the data; Then, statistical analysis based on fan ring neighborhood is used to remove outliers from the data and perform a second filtering on the data. Finally, multiple sets of 2D data after two rounds of filtering are restored to 3D point cloud data. Obtained a filtering accuracy of 97.8%, with a stable calculation time within 0.5s. The comparison of statistical filtering, radius filtering, and dynamic radius filtering algorithms shows that while improving the accuracy of laser point cloud filtering and the structured features of point clouds, it reduces running time and effectively suppresses dust and fog noise in point clouds.</p> <p>Aiming at the problems of threshold dependence and uneven density in the point cloud clustering of LiDAR, a point cloud clustering algorithm based on the reference plane projection of monocular camera and LiDAR data fusion was proposed. First, the enhanced image was edge detected using gradient quantile double threshold to obtain binary edge images; Secondly, the RANSAC method is used to extract the coal wall region from the filtered point cloud; Then, based on the calibration matrix of the monocular camera and LiDAR, the edge image is projected onto the point cloud data, and the coal wall edge is refined to obtain the coal wall point cloud with image edge constraints; Finally, the reference plane projection transformation method is used to cluster non coal wall point clouds and obtain point cloud clusters. The ground segmentation accuracy of this algorithm is about 93.5%, and the average clustering time is 58ms. Compared with the Kmeans and A-DBSCAN algorithms, the clustering results show an average decrease of 7.25% in DB index and an average increase of 9.06% in CH index. The results indicate that it can effectively reduce the phenomenon of over segmentation in point cloud segmentation, improve the inter class differences of clustered point cloud clusters, and reduce the intra class differences of point cloud clusters.</p> <p>A light intensity gray correction point cloud registration algorithm based on point cloud clusters is proposed to address the problem of low accuracy and high error matching rate in the registration of laser radar point clouds in fully mechanized mining faces. Firstly, calculate the feature values of the point cloud cluster and obtain the first set of candidate matching point pairs through feature matching; Secondly, extract key points and use extended feature matching to obtain a second set of candidate matching point pairs; Then, based on the reflected light intensity and grayscale of the laser, correct candidate matching point pairs, classify and process them, eliminate matching point pairs with larger distances, retain matching point pairs with smaller distances, and adjust the matching point pairs in the middle of the distance; Finally, the fish school search algorithm is used to solve the transformation matrix. The calculation time is stable within 0.35s, and compared with GP-ICP and CPD algorithms, the average RMSE decreases by 11.4%. The results indicate that it can effectively reduce the mismatch rate of point clouds and improve the accuracy of point cloud registration.</p> <p>Using a developed monocular camera and LiDAR data fusion system, combined with the proposed algorithm, on-site experiments were conducted on the fully mechanized mining working face at the Comprehensive Training Center of the Coal Major at Xi&#39;an University of Science and Technology. The theory and method studied in this paper were validated through the splicing experiment of the head point cloud, middle point cloud, and tail point cloud of the fully mechanized mining face, and good results were achieved. The point cloud registration method proposed in this article provides technical support for the transparent working face, and also established a theoretical foundation for achieving unmanned fully mechanized coal mining face.</p>