A novel automated method for doing registration and 3D reconstruction from multi-modal RGB/IR image sequences

摘要

In recent years, the use of multi-modal camera rigs consisting of an RGB sensor and an infrared (IR) sensor have become increasingly popular for use in surveillance and robotics applications. The advantages of using multi-modal camera rigs include improved foreground/background segmentation, wider range of lighting conditions under which the system works, and richer information (e.g. visible light and heat signature) for target identification. However, the traditional computer vision method of mapping pairs of images using pixel intensities or image features is often not possible with an RGB/IR image pair. We introduce a novel method to overcome the lack of common features in RGB/IR image pairs by using a variational methods optimization algorithm to map the optical flow fields computed from different wavelength images. This results in the alignment of the flow fields, which in turn produce correspondences similar to those found in a stereo RGB/RGB camera rig using pixel intensities or image features. In addition to aligning the different wavelength images, these correspondences are used to generate dense disparity and depth maps. We obtain accuracies similar to other multi-modal image alignment methodologies as long as the scene contains sufficient depth variations, although a direct comparison is not possible because of the lack of standard image sets from moving multi-modal camera rigs. We test our method on synthetic optical flow fields and on real image sequences that we created with a multi-modal binocular stereo RGB/IR camera rig. We determine our method's accuracy by comparing against a ground truth.