Comparison between Multi Epipolar Geometry & Affine 2d Transformation-Based Filters for Optical Robot Navigation

Abstract

Abstract: In previous research papers, Multi Epipolar Geometry-based Filter (M-EGF) has proven its high capability to overcome Single Epipolar Geometry-based Filter (S-EGF) as well as with conformal 2D Transformation-based Filter (C-2DF) in terms of providing precise, trusted, outlier-free, and real time Automatic Image Matching (AIM) results for Optical Robot Navigation (ORN) applications. This paper comes in a series of comparing M-EGF with the most common filters extensively used in ORN. Affine 2D Transformation-based Filter (A-2DF) is another familiar filter used for detecting outliers in AIM results and regards as the advanced version of C-2DF, where it can deal with any changing or distortion in image scales in X and Y directions. In this paper, M-EGF has been compared with A-2DF using the same system, tests, tracks and images and the same evaluation techniques used for comparing M-EGF to S-EGF and C-2DF in the related research papers. Tests show that A-2DF is similar to C-2DF in terms of its disability to filter AIM results in areas with open, narrow, and confused Depth Of Field (DOF). A-2DF is also limited in terms of finding out the 6 correct parameters for its mathematical model when AIM results includes a significant number of mismatching points. Difficult view angles has also a noted effect on the performance of A-2DF, but less than that of C-2DF due to its ability to work with different scales. In tests including limited DOF and low level of outliers, A-2DF has shown comparatively adequate outcomes, that is suitable for non-sensitive applications, in which the error does not entail any serious consequences. A-2DF is independent of the errors in Exterior Orientation Parameters (EOP) and Interior Orientation Elements (IOE) of cameras, where it is image points dependent estimation filter. Tests show that M-EGF is time- effective, efficient with any AIM findings, regardless DOF, capturing angle, outliers rate in observations, type of features and tracks. Unlike, A-2DF, tests show that the performance of M-EGF can affect by the quality of IOEs and EOPs, as a number of correctly matched points might be rejected, and this limitation is not related to the filter mathematical design but the professionalism of cameras calibration and EOP determination. in terms of processing time, A-2DF is considerably slow comparing with M-EGF due to its iterative and repeatability design, which makes it, unlike MEGF, unsuitable for real-time precise ORN applications. Keywords: Optical Robot Navigation, Epipolar Geometry, Automatic Image Matching, Co-planarity Equation, Real Time Applications, Affine 2D Transformation