Three-dimensional texture measurement using deep learning and multi-view pavement images

Surface macro-texture measurement plays an essential role in assessing pavement performance and quality. This paper presents a framework for macro-texture measurement and reconstruction using any number of pavement views. In the framework, an encoder based on deep CNN is used to extract features from single view pavement images. The features are then converted into a voxel model using a feature mapping unit and a decoder. A multi-view combination module aggregates the voxel models from any number of views to build a 3D model of pavement macro-texture. The 3D model can be used to assess pavement performance. The experiment demonstrates the effectiveness of the framework using a database with 16 asphalt pavements. The results show that the proposed framework performs 3D model reconstruction with satisfactory precision and stability. The 3D models assess mean texture depth and dynamic fraction coefficient with the average errors of 7.62% and 6.32%, respectively.

Automated summary

This study proposes a method based on deep learning to measure and reconstruct macro-texture using any number of pavement views, and demonstrates its effectiveness and stability through experiments. The 3D model allows for the assessment of pavement performance, with average errors of 7.62% for mean texture depth and 6.32% for dynamic fraction coefficient.