Vision-Based Roadway Geometry Computation

Abstract

Geometric data in transportation, such as roadway geometry, are important for asset management and for safety analysis. Traditional roadway geometric data are measured in the field, which is time consuming, costly, and dangerous. This paper proposes an algorithm to compute roadway geometric data, including roadway length, lane width, line width, and pavement marking size, from images. This paper makes two major contributions. First, the paper proposes a generalized roadway geometry computation algorithm using emerging vision technology based on two-dimensional (2D)/three-dimensional (3D) image reconstruction. The proposed algorithm consists of four steps, which are camera calibration from vanishing points, roadway vanishing line computation, homography computation and 2D/3D reconstruction, and, finally, roadway geometry computation. Second, the paper develops an error model, called roadway geometry error model (RGEM), to spatially quantify and visualize computation errors so that decision makers can choose measurement locations with an acceptable error. The geometric interpretation to RGEM is also presented in terms of roadway vanishing line. The proposed algorithm has been tested using two sets of images that were collected from the Georgia Tech campus and from actual video log images provided by the Georgia Department of Transportation. The roadway geometry was computed and the computation errors were analyzed. The test results show that the computation errors increase when the computation locations approach the roadway vanishing line. For the computation location with a distance of 190 pixels to the roadway vanishing line, the pavement lane width computation error is less than 3 cm. The experimental results also demonstrate that the proposed error model, RGEM, is able to reliably evaluate the roadway geometry computation errors. Applications of the proposed algorithm for modern and intelligent transportation system are also discussed.