Active learning of local structures from Attentive and Multi-Resolution Vision.
Maxime Cottret, Michel Devy · IAS · 2006
Abstract. In order to execute tasks and to na vig ate in an indoor en vironment, anautonomous mobile robot needs a comple x visual system to cope with detection,characterization and recognition of places and objects. W e are interested here in thede velopment of detection and characterization functions, inte grated on a compan-ion robot equipped with an omnidirectional color camera and a PTZ camera withpan, tiltand zoom controls. When learning a symbolic model of his w ork space, ourrobot executes concurrently a pre-attenti ve process and an attenti ve one; the formerone builds a map in which regions of interest (R OI) extracted from omnidirectionalimages (External localization and na vig ation modules are used in this case) arecoarsely localized. The latter analyzes more accurately each R OI by focusing thePTZ camera, grabbing a succession of vie wpoints and generating an appearancebased model of the region. Based on discriminant and invariant patches, this modelpro vides a set of visual features which will be used to cate gorize region classes andto recognize region instances in ne xt images.K eyw ords. local structures, attention, active vision, multi-resolution.