A Novel Low-Cost Double U-Net Model for Predicting Traffic Sign Retro-Intensity from Camera Data

Retroreflectivity is essential for the visibility of transportation infrastructure, ensuring road safety, especially under low-light conditions. Traditional methods for measuring retroreflectivity, such as nighttime visual inspections and retroreflectometer measurements, are labor-intensive, subjective, and pose safety risks. With the introduction of lidar technology, traffic sign retroreflectivity can be assessed more efficiently, as lidar-derived reflectivity values demonstrate a strong linear correlation with retroreflectivity. This study leverages a lidar device to propose a Double U-Net framework for predicting pixel-level reflectivity from daytime red, green, blue (RGB) images, providing a localized and accurate prediction. To train the Double U-Net model, a structured data set of over 7,600 images of transportation infrastructure was created, incorporating lidar-derived depth and reflectivity data. Given the sparsity of low-resolution lidar point clouds, linear interpolation was applied to generate pixel-level depth and reflectivity images. The proposed Double U-Net framework employs a two-stage architecture, where depth is predicted from cropped images in the first stage, and then combined with the original image and class embeddings in the second stage to generate pixel-level reflectivity predictions. A weighted loss function balances depth and reflectivity errors, enhancing prediction accuracy and robustness. The model achieved a median mean square error (MSE) of 0.0162 with interpolated data, 0.02233 with raw data, a median structural similarity index measure (SSIM) of 0.5413, and a Mann-Whitney U Test alignment of 58.2% with raw reflectivity data at a 0.001 significance level. The model effectively captures localized defects on traffic signs, providing a more detailed analysis compared with traditional methods.