Comparison Between Multi Epipolar Geometry and Conformal 2d Transformation-Based Filters for Optical Robot Navigation

Abstract

Abstract: This research paper comes in a series of comparisons between the Multi Epipolar Geometry-based Filter (M-EGF) and the common filters that are extensively used in Optical Robot Navigation (ORN) for filtering the results of Automatic Image Matching (AIM). The accuracy of ORN depends mainly on the quality of the AIM results. Conformal 2D transformation-based Filter (C-2DF) is a common filter used in ORN applications. However, C-2DF is limited in terms of: time processing, disability to deal AIM data with significant number of outliers, dealing with images taking from different Depths Of Fields (DOF) and difficult view angles. M-EGF has been introduced recently by the author and compared with Single Epipolar Geometry based Filter (S-EGF), showing high ability to provide ORN with precise, trusted, outlier-free, and real time observations. In this paper, M-EGF has been compared with C-2DF using images captured by optical navigation system simulating ORN, which includes 3 cameras synchronized using GPS time. The performance of the two filters has been evaluated in different AIM environments and the automatically filtered results have been compared to precisely manually reviewed matched points using Matlap. Tests show that C-2DF has failed to deal with the AIM results in areas with open, narrow, and confused DOF. Also, they have failed to find out the right mathematical model in data with high rate of mismatched points and images with difficult view angles. C-2DF is also limited in its capability to deal with figures including different scales. With limited DOF and limited rate of errors, C-2DF has provided relatively sufficient results, which can be suitable for ORN in terms of quality and processing time. C-2DF overcomes S-EGF and M-EGF in terms of being not affected by errors in the Exterior Orientation Parameters (EOP) and Interior Orientation Elements (IOE) of cameras, as C-2DF depends on estimating the mathematical model parameters using only image points.