Friction-ResNets: Deep Residual Network Architecture for Pavement Skid Resistance Evaluation

Abstract

Pavement friction and surface texture are crucial for highway safety. Acknowledging the gaps in understanding the relationship between pavement surface friction and texture, this paper introduces a novel deep residual network (ResNets), named Friction-ResNets, tailored for pavement friction prediction based on surface texture data sets. The Friction-ResNets architecture consists of 11 convolution layers, 1 average pooling layer, and 1 fully-connected layer with millions of neurons. Different from deep convolutional neural networks (CNNs), Friction-ResNets are used as a residual learning framework with skip connections to significantly lower gradients and enable the effective training of much deeper networks for improved classification accuracy. There are 33,600 pairs of friction and their corresponding texture data was collected and prepared from multiple pavement surface types distributed in 12 states for training, validating, and testing of Friction-ResNets. The testing results show that Friction-ResNets can achieve a classification accuracy of 91.3%, outperforming the four conventional machine learning methods (Gaussian Naïve Bayes, k-nearest neighbors, support vector machines, and random forests) investigated in this study by a wide margin. The application of ResNets in this study demonstrates the potential of using highway speed noncontact texture measurements for pavement friction evaluation using deep learning algorithms.