计算机视觉技术是通过结合摄像机的捕获功能和计算机的处理能力来模拟人眼，进而对视频帧中的目标进行识别、跟踪和测量的机器视觉技术。对井下作业进行视频监控可以提高井下工作的管理水平和安全性。在井巷工程中，挖掘巷道后必须对岩石进行加固，锚杆支护是常用的加固方式，锚杆支护质量直接关系到井下工作的安全。锚杆支护安装质量评判标准包括锚杆的材质，锚杆钻孔深度与锚杆间排距等，而锚杆钻孔深度是影响锚杆支护质量的一个重要因素，保证钻孔深度必须保证钻入锚杆的数量。 利用计算机视觉技术完成锚杆计数，进而保证锚杆支护质量是本文的研究课题，为了实现统计计数必须对锚杆机进行监控跟踪。 井下视频图像缺乏彩色信息，且光照环境复杂。针对具体的井下锚杆机跟踪问题，在研究多种跟踪算法的基础上，并结合具体环境采用基于TLD的目标跟踪算法。本文主要工作内容如下： (1) 采用基于稀疏光流的Lucas-Kanade金字塔光流算法对锚杆进行跟踪。在跟踪点的选择方式上采用均匀选点和角点相结合的选取策略。通过计算相邻前后两帧的双向跟踪结果的误差，以及根据对应跟踪点周围区域的匹配相似度并与阈值进行比较来对跟踪结果点集进行过滤。 (2) 采用基于分类的方法进行目标检测。目标的搜索策略上采用局部搜索与全局搜索相结合的方式，上帧成功跟踪时，进行目标的局部搜索，上帧跟踪失败时进行全局搜索。采用具有全局信息的方差阈值分类，基于图像特征的快速随机森林分类和具有高可靠性的利用模板匹配相结合的方法进行分类检索目标。 (3)采用PN学习方法对检测的分类器进行学习更新。跟踪失败时采用线性预测方法实现对目标位置的预测。根据跟踪历史信息实现对锚杆机运动轨迹的绘制，并完成数量统计工作。采用Openmp并行指导实现对跟踪和检测的并行化处理，提高跟踪速度。 实验结果表明，采用TLD跟踪算法可以实现对锚杆机的跟踪，完成锚杆数量的计数工作，并经过改进后提高了TLD算法的性能。

Computer vision technology is the technology to detect, track and measure the target in video through combining the capture function of the camera and processing function of the computer. The video surveillance can improve the safety and management level of the underground work. Among all the work in mine, we must strengthen the rock after mining the tunnel. Bolt supporting is the common way to strengthen the rock, and the quality of bolt supporting is related to the safety of the work underground. The evaluation of the supporting quality includes the bolt material, the depth of the bolting hole and the distance between the bolts. The depth of the bolting hole is an important factor affecting the quality of the supporting. In order to ensure the depth of the bolting hole, we must ensure the count of the bolts. Counting the bolts through the computer vision technology is the subject of the paper. We must track the roof bolter in order to count the bolts. The video in mine is lacking of color information and the illumination is complex. Having researched several tracking algorithms, we adopt the TLD algorithm to track the roof bolter in mine based on the special issue. The main contents in our paper as follows. (1) Adopting the Lucas-Kanade algorithm based on the sparse optical flow to track the roof bolter. Selecting the corners and uniform points as the tracking points. To filter the result points, we calculate the forward errors between two adjacent image frames and the matching similarity of the area around the result points, and compare them with the threshold. (2) Adopting the classification method to detect the target. The combination of the global search and local search is adopted to detect the target, if the tracking is successful in last frame, we use the local search, otherwise, we use the global search. The variance that contains global information, the random forest based on image features and the template matching method which has high reliability are used to detect the target through classing. (3) The PN learning method is adopted to update the classifier. Adopt the linear predicting method to predict the location of the target when the algorithm fails. We draws the trajectory of the roof bolter according to the historical information, and complete the task of the counting. To increase the speed of the algorithm, we use the Openmp library to realize the parallel processing of tracking and detection. The results shows that the improved TLD algorithm in our paper can track the roof bolter, count the bolts , and we can get better performance using the improved algorithm.