A New Sense for Depth of Field

Abstract

One of t he major un so lv ed pro bl ems in designing an autonomous agent jr obotj th at mu st function in a co mplex, moving environment is obtain in g re li able, real-time depth information, preferahly without the li mi tations of ac ti ve sc anners. Stereo remains computat ionally intensive a nd prone to seve re errors, the use of motion information is still quite experim en ta l, a nd autofocus schemes can measure depth at o nl y one p oi nt at a t im e. We examine a novel source or depth information: focal gradie nts res ul t ing fr om t he limited depth or fleld inherent in most optical system s. We prove t hat this source of information can be use d to make re li a bl e depth maps of useful accuracy with relatively minima l computatio n. Ex periments with realistic imagery show that measu rement or these optica l gra di ents c an potentially provide depth inform at ion rough ly co mparable to stereo disparity or motion parallax, whil e avoidin g im age-to-image mat ehing problems. A potentially real-time ve rsion of t hi s algorithm is described.