École Nationale Supérieure de
Techniques Avancées

Abstract

A burning issue in autonomous robotics is to build an environment representation and to let the robot localize itself inside this representation. Simultaneous Localisation And Mapping (SLAM) are computing these two asks. The Cognitive Robotics of the french Ecole Nationale Supérieure des Techniques Avancées (ENSTA, Paris), directed by David Filliat, developed a rapid, qualitative, discontinuous, incremental and purely visual SLAM. This system was build on sony Aibo robots and uses “bag of visual words” methods and a voting algorithm to perform scene-recognition. Performance is about 2 seconds per recognition task, for a recognition rate of 80%. The work presently reported was about introducing geometrical information in the regognition system. This goal was reached using oriented segments of two visual words as new voting objects. The combinatorial issue is controled, and computing time only increased by less than 2. Geometry introduced strong robustness to vocabulary parameters. Implemented solution is quite rotation invariant and independent to the word vocabulary used (both SIFT and color histograms where tested). While not realy inproving system performance, this work brings best understanding of parameters impact on the vision system created, and opens new hopeful research axes.