Detecting Pavement Joints and Grooves with Improved 3D Shadow Modeling

Abstract

This paper proposes a modification of the original three-dimensional (3D) shadow modeling for improvements in suppressing noises. Compared with the original 3D shadow modeling (3D Shadow Modeling I), this improved 3D shadow modeling (3D Shadow Modeling II) presents algorithmic improvements to eliminate noises by inspecting the general lighting behaviors at the specified local neighborhood centered at each individual pixel. In 3D Shadow Modeling II, bidirectional lighting and local neighborhood inspection, which are the two fundamental procedures of 3D Shadow Modeling II, are repeated with various rotation angles to detect descended patterns of arbitrary orientations. With respect to the detection of pavement joints and grooves, the proposed 3D Shadow Modeling II inspects the general lighting behaviors at a long one-dimensional (1D) local neighborhood centered at each image pixel. Through inspections at sufficiently long 1D local neighborhoods, 3D Shadow Modeling II is capable of eliminating noises more efficiently and does not require any subsequent procedures to further remove noises in most cases. To support intensive computations resulting from the long length of the 1D local neighborhood, a massively parallel computing method using a general-purpose graphics processing unit (GPU) is developed in the paper for improvement of computational efficiency at two orders of magnitude. According to the experiments on 2 3D testing images, 3D Shadow Modeling II can produce consistently high precisions and recalls over 9% for almost all testing images. The overall precision and recall for testing images with joints are 98.58 and 97.46% respectively, while the overall precision and recall for testing images with grooves are 96.49 and 93.87% respectively. It is anticipated that further research and development will demonstrate that the proposed 3D Shadow Modeling II would be more competitive than 3D Shadow Modeling I in detecting other descended patterns such as cracks and potholes, once the local neighborhood for inspection is defined properly to take advantage of the geometric features of the target descended patterns.