A non-feature based method for automatic image registration relying on depth-dependent planar projective transformations

摘要

Multisensory data fusion oriented to image-based application improves the accuracy, quality and availability of the data, and consequently, the performance of robotic systems, by means of combining the information of a scene acquired from multiple and different sources into a unified representation of the 3D world scene, which is more enlightening and enriching for the subsequent image processing, improving either the reliability by using the redundant information, or the capability by taking advantage of complementary information. Image registration is one of the most relevant steps in image fusion techniques. This procedure aims the geometrical alignment of two or more images. Normally, this process relies on feature-matching techniques, which is a drawback for combining sensors that are not able to deliver common features. For instance, in the combination of ToF and RGB cameras, the robust feature-matching is not reliable. Typically, the fusion of these two sensors has been addressed from the computation of the cameras calibration parameters for coordinate transformation between them. As a result, a low resolution colour depth map is provided. For improving the resolution of these maps and reducing the loss of colour information, extrapolation techniques are adopted. A crucial issue for computing high quality and accurate dense maps is the presence of noise in the depth measurement from the ToF camera, which is normally reduced by means of sensor calibration and filtering techniques. However, the filtering methods, implemented for the data extrapolation and denoising, usually over-smooth the data, reducing consequently the accuracy of the registration procedure...