Attentive Strategies for Autonomous Navigation

摘要

This paper describes a "dynamic" approach to vision-based navigation in unstructured environments. The navigation system combines the information derived from internal odometry and from images collected by a single camera mounted on-board. The software/hardware architecture presented in this paper is based on the automatic collection of images and the organization of visual features into visual graphs, covering the whole path. The control of the vehicle is based on the visual servoing paradigm and it is performed by continuously and incrementally updating the vehicle's trajectory. The automatic selection of generic visual features of the scene is performed by a newly implemented multi-resolution technique, based on morphological filtering and a "winner take all" strategy. Even though this technique allows to find any type of generic feature, in our implementation we restrict the analysis to corners. The planning module is based on the recognition of the current image within the current image mosaic of the visual graph. Results related to the navigation of a mobile robot along simple indoor paths are presented.