Deep-learning based non-contact method for assessing pavement skid resistance using 3D laser imaging technology
Guangwei Yang, Allen Zhang, Kelvin C. P. Wang, Qiang Li, Wenyao Liu, Yang Liu
Abstract
Traditional contact and water-based testing devices exhibit some limitations in evaluating pavement skid resistance. This study demonstrates the viability of deep learning (DL) based non-contact pavement skid resistance evaluation using 3D laser imaging technology. Pavement 3D images and friction numbers were collected from 28 field sites in 10 states in the U.S. to cover pavement sections with various texture and friction characteristics. A total number of 28,000 pairs of 3D images and friction numbers were prepared as the database to train and validate the friction prediction model, named as FrictionNet-V, using the convolutional neural network. The input of FrictionNet-V is the 3D pavement image from the left wheel path, and the output is the corresponding friction number ranging from 0.2 to 0.9. Based on the experimental results on 5,600 testing images, FrictionNet-V achieved better performance than other DL models in terms of speed and accuracy. The result demonstrates the potential of highway speed non-contact pavement skid resistance evaluation using high resolution 3D texture images and DL methods at the network level in the future.