A novel approach for pavement distress detection and quantification using RGB-D camera and deep learning algorithm

Abstract

This study presents an innovative and cost-effective approach for pavement distress detection and quantification using an RGB-D camera in conjunction with an instance segmentation algorithm. The proposed method employs the instance segmentation algorithm to acquire 2D coordinate information pertaining to the affected pixels and integrates the internal reference matrix and depth data from the depth camera to transform the 2D images of the distressed areas into 3D point cloud data, consequently enabling distress quantification. To assess the dependability of the proposed approach, high-density polystyrene foam was used to simulate potholes and cracks of the pavement. Data collection was conducted under varying experimental conditions to identify the optimal data collection scheme. Using the approach, potholes and cracks of varying severity were collected from both asphalt and cement concrete pavements. The outcomes of the study demonstrate that the proposed methodology is capable of accurately detecting and quantifying potholes on the pavement. Furthermore, the errors associated with both the calculated area and volume exhibit a gradual decrease with an increase in the severity of pavement distress. However, for cracks, the method yields a larger error in results, primarily due to the instance segmentation algorithm's imprecise segmentation of crack edge pixels, despite superior restoration of the planar contour of cracks.