Automated two-dimensional–three-dimensional registration using intensity gradients for three-dimensional reconstruction

Abstract

We develop a robust framework for the registration of light detection and ranging (LiDAR) images with 2-D visual images using a method based on intensity gradients. Our proposed algorithm consists of two steps. In the first step, we extract lines from the digital surface model (DSM) given by the LiDAR image, then we use intensity gradients to register the extracted lines from the LiDAR image onto the visual image to roughly estimate the extrinsic parameters of the calibrated camera. In our approach, we overcome some of the limitations of 3-D reconstruction methods based on the matching of features between the two images. Our algorithm achieves an accuracy for the camera pose recovery of about 98% for the synthetic images tested, and an accuracy of about 95% for the real-world images we tested, which were from the downtown New Orleans area.

Automatic Registration of Panoramic Image and Point Cloud Based on the Shape of the Overall Ground Object

“…The region-based registration method by converting the point cloud into an intensity or distance image, and then comparing the pixel similarity between the point cloud image and the optical image for registration. This comparison can be done using various metrics such as gradient information [18], phase correlation [19], mutual information [5,20,21], normalized joint mutual information [22], joint entropy [23], etc. Unlike featurebased registration methods, the region-based approach does not require feature extraction from the data and instead compares the data's correlation in the corresponding regions.…”