Game engine-driven synthetic point cloud generation method for LiDAR-based defect detection in sewers

Abstract

3D point clouds acquired from emerging light detection and ranging (LiDAR) enable quantitative defect inspection for urban sewers. Defect detection of sewer point clouds is an essential step in deep-learning-based condition assessment; however, it faces challenges related to data scarcity. This study proposes a novel LiDAR simulation-based synthetic point-cloud data generation methodology driven by game engines, which addresses issues arising from geometric occlusions and the defect randomization. Evaluation experiments were conducted to validate the enhancement provided by the synthetic point clouds. The results indicated that the synthetic data significantly enhanced the accuracy of detecting corrosion and spalling defects by 17.90% and 4.69%, 15.22% and 11.69%, 15.56% and 15.21%, and 10.32% and 1.40% for 50, 100, 200, and 300 real scans, respectively. In addition, pre-trained models outperformed the widely used hybrid training approach, solving knowledge transfer between synthetic and real data. The proposed method alleviates data scarcity and contributes to automated defect segmentation in sewer point clouds.