Translation Estimation in Log-Polar Images of Onboad Cameras and Generating Night Road Orthoimages using Phase-Only Correlation

Abstract

This study investigates a method for generating road orthoimages using onboard images taken at night by a high-sensitivity digital camera for avoiding the effects of sunlight and human/vehicle congestion. Perspective projection images from onboard cameras are generally transformed into overhead images by projection transformation using feature quantities such as feature points. The mosaicing process is also based on the feature quantities between the overhead images. Therefore, the generation of feature-based night road orthoimages is an ill-posed problem. Log-polar transform (LPT) with rotation and scale invariance and the correlation-based phase-only correlation (POC) method are used for the above problem. However, in order to keep the properties of the LPT, the problem of translation in Log-polar images has to be resolved.

With this motivation, simulated images in which the reference image is moved slightly in the direction of the car's travel sequentially generate and determine the simulated-reference image under the similarity between each simulated image and the target image. Furthermore, the reference image is geometrically corrected by the LPT between the simulated-reference image and the target image, and the generation of night road orthoimages as an ill-posed problem becomes a well-posed problem by the POC on the corrected reference image and the target image is described in this paper.