(General session) Mining and underground construction machineries / Rock Engineering / Mining technologies

4:20 PM - 4:40 PM JST(7:20 AM - 7:40 AM UTC)

[2416] UAV-Based 3D Gaussian Splatting for Rapid Reconstruction and Visualization of Underground Tunnels

○Adrian Fungayi Binala^1,2, Narihiro Owada², Hisatoshi Toriya², Tsuyoshi Adachi² (1. Student, 2. Akita University)

Chairperson: 福田大祐（北海道大学）

Keywords:

Unmanned Aerial Vehicles (UAVs),3D Gaussian Splatting,Cube mapping

Unmanned Aerial Vehicles (UAVs) have gained unparalleled importance in engineering applications. Their use has increased significantly in various fields of engineering owing to their noninvasiveness and versatility. Non-invasive mapping can be performed using drones. In this study, we examined the use of 3D Gaussian Splatting for the quick reconstruction of subterranean tunnels employing UAVs in deep, dark, GPS-denied environments, such as mining and civil engineering. Conventional 3D modeling techniques have high computational demands and poor texture under low-light conditions. To address these issues, 3D Gaussian splatting was implemented in this study. To evaluate this technique, a drone equipped with a 360-degree camera and artificial illumination was sent into an underground tunnel to collect video data. Frames were extracted, cube-mapped, and reconstructed into five distinct 3D models with different splat counts for comparison. The results demonstrated high visual fidelity, with the SSIM improving from 0.649 to 0.716. Initial geometric accuracy revealed a limitation, with the RMSE remaining at approximately 1.8 m across all training stages. However, after applying scale correction and geometric adjustment in CloudCompare using known reference measurements, the accuracy of the reconstructed model improved significantly, achieving an overall RMSE of 0.05366 m, with directional RMSE values of 0.03195, 0.06548, and 0.05460 m in the X, Y, and Z directions, respectively. UAV-based 3D Gaussian splatting is an effective tool for rapid 3D visual inspection and real-time rendering, demonstrating reliable geometric accuracy after scaling correction and strong applicability in engineering operations.