Visualization and intelligent solutions for big pavement data

Abstract

Pavement visualization and crack detection are two important components supporting modern pavement condition survey. In this dissertation, two major goals are accomplished based on implementable algorithms. 1) The long-distance pavement 3D visualization is developed based on an effective Level-of-Details (LOD) algorithm named "Geometry Clipmap". The Geometry Clipmap is able to render large-scale 3D scene by caching pavement height data in a set of nested grids. During 3D rendering process based on Geometry Clipmap, the data size uploaded to video memory is reduced significantly, while the detailed high-resolution pavement surface data are still retained for further detailed inspection. As a consequence, Geometry Clipmap provides long-distance pavement display with excellent visual continuity and stable frames per second (FPS). In addition, the size of large-area surface distresses can be directly measured under long-distance pavement 3D environment. This is the first study that applies LOD algorithm to long-distance 3D pavement visualization and large-scale distress measurement. 2) The pixel-level automated crack detection is achieved through the modifications of Cell-based Convolutional Neural Network (CNN) on 3D pavement surface. CNN is a deep learning algorithm that targets at recognizing Cell-based imaging objects. Based on combination of specific pre-process layers and new gradient computational strategy, the network for pixel-level detection is developed in the study for effective feature learning. In addition, consecutive convolutional layers and a new activation unit are introduced to improve the detection performance at fine and shallow cracks. Furthermore, GPU parallel computing techniques are utilized to speed up the proposed pixel-level detection network. As a result, the CNN based pixel-level crack detection outperforms traditional imaging algorithms in terms of F-Measure. Lastly, the highly-efficient Deep-Learning network for speedy processing fits well for big pavement datasets processing.