Steps Towards One-Shot Vision-Based Self-Localization

Summary

Approaches to the problem of self-localization are based on odometry (dead-reckoning) and the matching of feature vectors obtained from sensor readings, captured at selected positions and orientations (landmarks) during map construction. Due to the big uncertainties of sensor inputs, the matching of their readings is taken as a correction for odometry-based self-localization. Vision-based self-localization is based on the recognition of landmarks or the matching of 3D information extracted from the images. Such approaches are computationally expensive. We propose the use of Heteroassociative Morphological Memories for the fast recognition of views that can be used in a one-shot recognition procedure, inspired in the instantaneous recognition of specific views of an environment where an autonomous agent tries to obtain orientation. Recognition of world views provide inmediate confirmation or refutation of the position hypithesis that other senses or the dead-reckoning procedures may sustain. The present status of our work addresses the problem of significative shot identification.