Boundary detection places an important role in underground coal mine autonomous driving technologies. Accurate boundary information is beneficial to object detection and subsequent decision and control of autonomous trackless rubber-tyred vehicles. To meet the urgent demands in underground boundary detection, a method for underground boundary detection and tracking is proposed. The proposed method consists of four parts: real-time point cloud correction, point cloud gridding, boundary fitting and boundary tracking. Based on the characteristics of irregular boundaries and frequently changing slopes in underground mines, a real-time point cloud correction method is designed, which consists of pre-correction and dynamic correction to achieve parallelism between the road point cloud and the LiDAR coordinate system. To avoid the influence of uneven road surfaces on boundary extraction, the corrected point cloud is projected into a two-dimensional gridding map. To address the problem of blind areas caused by turnings and chambers, the Kalman Filter is utilized to track the missing point cloud. The performance of the proposed was evaluated on the data collected in real-world operation scenarios. The experimental results demonstrate that the proposed method can detect and track underground boundaries with the precision of 91.2%, as well as meet the real-time requirements in autonomous driving.