Multi RobotObject Tracking andSelf Localization UsingVisual Percept Relations
Daniel Göhring · 2006
Inthispaperwe present a novelapproach to estimating theposition ofobjects tracked byateamofmobile robots andtousetheseobjects fora better self localization. Modeling ofmovingobjects iscommonly doneinarobo-centric coordinate framebecause this information issufficient formost lowlevel robotcontrol anditisindependent ofthequality of thecurrent robotlocalization. Formultiple robots tocooperate andshareinformation, though, theyneedtoagreeonaglobal, allocentric frameofreference. Whentransforming theegocentric object modelinto aglobal one,itinherits thelocalization errorof therobotinaddition totheerrorassociated withtheegocentric model. We propose using therelation ofobjects detected incamera images tootherobjects inthesamecameraimageasa basis forestimating theposition oftheobject inaglobal coordinate system. Thespacial relation ofobjects withrespect tostationary objects (e.g., landmarks) offers several advantages: a)Errors in feature detection arecorrelated andnotassumedindependent. Furthermore, theerrorofrelative positions ofobjects within a single cameraframeiscomparably small. b)Theinformation isindependent ofrobotlocalization andodometry. c)As a consequence oftheabove, itprovides ahighly efficient method forcommunicating information abouta tracked object and communication canbeasynchronous. d)Asthemodeledobject isindependent fromrobo-centric coordinates, itsposition canbe usedforself localization oftheobserving robot. We present experimental evidence thatshowshowtworobots areabletoinfer theposition ofanobject within aglobal frame ofreference, eventhough theyarenotlocalized themselves and thenusethis object information forself localization. IndexTerms-Sensor Fusion, Sensor Networks