Robust motion estimation using covariance intersection

Robust features extraction, matching and tracking are crucial for many important applications in many computer vision systems. The accuracy of features locations is depending on the variation in intensity within their neighbourhoods and from which their uncertainties are estimated. In this paper we present a technique to improve the accuracy and the robustness of the fundamental matrix and consequently the motion estimate by considering those uncertainties. In our solution, rather than converting colour images to gray level images, resulting in a huge loss of information as most vision applications do, each channel of the RGB images is processed separately. Then the covariance intersection technique is adopted to fuse all uncertainties in each channel. Through several experimental results in different environments, we show that by including the fused feature uncertainties from all three channels of RGB images, better estimates of the fundamental matrix are obtained.