Evaluation of the Accuracy of a 3D Temple Model Based on Simultaneous Localization and Mapping Technology (Case Study: Pura Agung Wira Loka Natha, Cimahi)

Abstract

Simultaneous Localization and Mapping (SLAM) technology rapidly generate 3D point cloud models of complex structures. This study evaluates the accuracy and efficiency of the 3D model of Pura Agung Wira Loka Natha generated using two SLAM scanning approaches: single-session closed-loop and multi-session scanning. The resulting point clouds were processed and validated using field measurements obtained from a total station, a measuring tape, and a laser distance meter. To assess dimensional accuracy, 18 building elements were distributed across Nista, Madya, and Utama Mandala zones. Positional and dimensional evaluations were performed using Root Mean Square Error (RMSE), Circular Error 90% (CE90), Linear Error 90% (LE90), and Paired t-Tests. The results show that the single-session approach achieved an aggregate dimensional RMSE of 0.00284 m, while the multi-session approach yielded 0.00364 m. The Paired t-Test produced a p-value of 0.671, indicating no statistically significant difference in dimensional accuracy between the two approaches. However, the multi-session approach produced a denser point cloud (74,008 points/m² vs 52,317 points/m²) due to its longer scan duration and high trajectory overlap, while achieving better horizontal positional accuracy (CE90 of 0.043 m vs 0.053 m) because dividing the survey into shorter, independent trajectories restricted horizontal odometry drift propagation. Conversely, the single-session approach achieved better vertical positional accuracy (LE90 of 0.013 m vs 0.074 m) by leveraging continuous loop-closure to globally optimize vertical drift and avoid sub-map merging residuals. In terms of efficiency, single-session scanning took 26 minutes and 11 seconds, whereas multi-session scanning required 40 minutes and 22 seconds. This study demonstrates that for time-constrained and vertical-accuracy-critical heritage documentation, the single-session approach is highly preferable due to its unified loop-closure stability, conversely, for documentation applications prioritizing dense point clouds and superior horizontal positional consistency for highly detailed ornamental representations, the multi-session strategy is recommended.