Pavement texture depth estimation using image-based multiscale features

Abstract

Texture depth, a fundamental indicator for pavement performance, is traditionally obtained by time-consuming measurements. The image-based estimation has become a new trend due to its convenience and economy. This study applies image-based multiscale features for texture depth estimation. Maximum particle size distribution (MPSD) and relative energy distribution (RED) are proposed based on multiscale segmentation and 2D-wavelet decomposition. Two hundred fifty image samples labelled with electronic mean texture depth (eMTD) were collected. The multivariable nonlinear regressors are developed to deal with features' multicollinearity. As a result, the models where the input is the combination of MPSD and RED have better performances than those where the input is two sets of features. The random forest model yields the best results (cross-fold validation R2 = 0.8192). The proposed method has the potential to enhance vision-based MTD measurements, which supports pavement quality evaluation during construction.