Rock segmentation visual system for assisting driving in TBM construction

The tunnel boring machine (TBM) is a key equipment for excavating long-range tunnels. It is a complex system and hard to be controlled well in practice. In this paper, we propose the rock segmentation visual system to assist TBM driving. Through the system, online size distribution of excavated rocks is automatically analysed and sent back to TBM driver, from which many statistical information can be gathered. The system's core algorithm is based on semantic segmentation, and the rock detection task is viewed as a rock/background pixel-wise classification problem. Accordingly, the Rock Segmentation Dataset is made with specific annotation strategies, and the goal of the dataset is to pick out large rocks in the images. Many networks are evaluated quantitatively on it, and we select the best suited one. We design two parallel networks to extract rock object and contour mask, such that the connected rock areas in object mask can be split with a mask fusion algorithm. Further network modification is made to boost inference speed that meets the requirement of system design. Experimental results show that the system can effectively detect large rock particles in the images and make necessary statistical analysis. Specifically, the segmentation accuracy achieves 68.3% mIoU, and the inference speed achieves 19.4 FPS under image resolution of 1600x1200on one NVIDIA Titan XP GPU. From the viewpoint of statistical analysis, 43.5% rock size IoU and 14.7% error rate of mean rock size are obtained, which is acceptable from the viewpoint of real applications.

Automated summary

The study introduces a rock segmentation visual system to assist the driver of a tunnel boring machine. The system can automatically analyze the size distribution of rocks and provide statistical information, effectively detecting large rock particles and meeting the speed requirements of the system design.