Unsupervised pavement rutting detection using structured light and area-based deep learning

Abstract

Timely and comprehensive pavement rutting detection is crucial for road safety and maintenance. Traditional methods often fail to capture full morphological characteristics and severity of rutting. This paper proposes an area-based pavement rutting detection method using unsupervised deep learning. An adaptive point cloud rasterization strategy and multi-feature mapping enhance surface detail preservation while reducing complexity. A deep learning model segments rutting based on feature similarity and spatial continuity, refined by point cloud reconstruction and post-processing. Tested on a 600 km roadway dataset with 706 rutting samples, the method achieves 91.46 % accuracy, surpassing conventional models. It maintains high efficiency, reduces labeled data reliance, and requires only structured light-based scanning, making it suitable for large-scale applications. Ablation studies validate the multi-feature fusion strategy, establishing a new paradigm for high-precision rutting detection. Successfully deployed in real-world inspections, this method advances infrastructure assessment within smart transportation systems.