A Variable Precision Hybrid Camera Calibration Method


We propose a hybrid camera calibration based on CVP (calibration using vanishing points) and BFC (brute-force calibration) that can produce solutions of varying precision. The application context that we are looking at requires matching a 3D model of known dimensions to a single image of the object. As such, our algorithm only requires a single image of a cuboid object (representing the bounding box) as input. CVP is a closed-form calibration method and provides a quick initial estimate. However, it gives larger errors when applied to low resolution images. BFC is an iterative solution that progressively provides higher accuracy with each successive refinement. But BFC, by itself, is very time consuming because it has to search through a large space of possible transformations of the model. We therefore propose to apply CVP first to obtain a coarse estimate of the transformation and then apply BFC to refine the solution. This can be viewed as either improving the CVP accuracy or speeding up BFC. In this paper, we also provide comparisons of the accuracy of four different calibration techniques: CVP, BFC, HC (hybrid calibration), and Tsai's methods. The results show that our hybrid method is the best one in general.


The IASTED 2001 paper is available.


Hough transform of cuboid
Detected corners
Detected edges
Detected edges

Return to AVIS home page.