Video-based localization without 3D mapping for the visually impaired

In this paper, we present a system for indoor human localization that does not need 3D reconstruction of features or landmarks. We assume that a video sequence has been acquired and that keyframes have been registered with respect to 2D positions and orientations. In online mode, we use only a handheld monochrome fisheye camera and a synchronized IMU as sensory inputs. The query is not based on a single image but uses a HMM-based state estimator. Our image representation consists of initial global GIST vectors followed by local SURF features. We present a novel approach to localization by using search space reduction on global features, then HMM based position prediction and estimation on local features. Experimental results show that accurate localization is achieved and realtime performance is feasible. This work demonstrates that a working portable system could be designed for the visually impaired.