講演情報
[Data_Collection-12]Efficient Prediction of Highway Friction from High-Speed Texture Profiles
*Can Chen1, Christian Andres Sabillon-Orellana2, Danilo Inoue1, Jorge A. Prozzi1 (1. University of Texas at Austin (USA), 2. University of Tennessee, Knoxville (USA))
キーワード:
surface texture、safety、friction prediction、deep learning、network-level measurements
A reliable model for evaluating pavement friction performance is crucial for effective pavement management and traffic safety. Existing models, however, face several challenges, including their dependence on laboratory test data (which is often unavailable at the network level). To address these limitations, this study proposes two models for predicting pavement friction from high-speed two-dimensional (2D) transverse surface profiles: a traditional regression model and a deep learning model. Using a host vehicle equipped with in-house laser-sensor hardware, high-speed 2D surface profiles were collected in the vicinity of Austin, Texas, along with skid number measurements from a Locked-wheel Tester, covering more than 150 km (93 mi) of texture and friction data. Because the raw texture profiles contain trends, spikes, flat signals, and other types of unwanted outliers, pre-processing steps were first applied to clean and standardize the data. The two friction models were then developed using distinct sets of input features: (i) seven uncorrelated texture parameters derived from processed 2D profiles for the regression model, and (ii) a comprehensive feature space of over 50 profile-based indices for the deep learning model. The analysis was conducted at the project level using 0.1-mile sections, but the proposed methodology can be generalized to network-level applications. While the models achieved accuracy levels of approximately 80 percent, their value lies not only in estimating friction values but also in identifying critical or low-friction “hot spots”. Such a tool could enhance pavement management decision-making, improve roadway safety, and enable efficient, large-scale friction assessment with accuracy comparable current technologies.
