Distortion-Free Navigation of Omni-Directional Images Using Constructive Neural Networks
Frank Edughom Ekpar, Masaaki Yoneda, Hiroyuki Hase · 2000
111 tliis paper, we present a novel approach to the generation of arbitrary perspective-corrected views from onlni-directional inlages for the purpose of interactivr navigation of tlie images in real time. We use a constrrlctivc nrnral network of suitable conlplexity to lrarri tlie inlierent distortion of the ornni-directional iniaging systc.ni llsiiig training sets obtained through carc~fi~lly constructed cali1)ration pattmns. Starting with a near-niiiiinial neural network, the topology of tlie xieural network is niodified al~t~oniatically over snccessivt ~ training cycles until a reasonable, nearoptinla1 xic.twork is ol)tained. Our system overconles the liriiitations of prc.vious methods by obviating the nt3eti to dclrive t,lle perspective projection equations of the oriiili-dirc>ctio~ial iinaging sjrsteni. Addition-ally, ollr systfc~ni is rol~ust enough to correctly approximate an omni-directional imaging system of arbitrary conlplexity yet elegantly simple enough to permit real-time corrcctiori for most omni-directional imaging systems currently available. We demonstrate the practicability of our approach by describing its ap plication to the generation of arbitrary perspectivecorrected views from fish-eye images. 1