Optimization-Based Methods for Road Image Registration

摘要

A number of transportation agencies are now relying on direct imaging for monitoring and cataloguing the state of their roadway systems. Images provide objective information to characterize the pavement as well as roadside hardware. The tasks of processing, interpreting, and assessing the condition of the pavement from image data sets poses formidable challenges however. Not only are the data sets extremely large but they are not taken under ideal conditions neither in geometric location and alignment, nor in photometric consistency due to a number of factors including the inaccuracy of dead reckoning as well as the dynamics of the collection vehicles. The resulting images provide, in essence, a set of independent overlapping snapshots of regions of the pavement that must be merged together to produce the continuous mosaic that is needed for automated analysis and interpretation. This requires the registration of these images. Registration the problem of aligning features in one image/view to corresponding features in another image/view of the same object. Registration brings independent images into the same reference frame. To be effective, this process must be automated requiring minimal or no user intervention. In this work, we describe a non-parametric optimization procedure that is capable of producing general nonlinear deformations to register multiple images so they can be merged, compared and analyzed.