Image Based Location Estimation for Walking Out of Visual Impaired Person

A new and intelligent walking navigation system could be helpful for visually impaired people so that they do not need helpers or guide dogs on going out. Conventional navigation systems using GPS are not available indoors or undergrounds. A new location estimation method that is available in such situations is necessary. We propose to estimate pedestrian's location with only a single camera attached to the pedestrian. Our proposed method uses only computer vision and no other sensors are needed. The location estimation is achieved by image retrieval. The retrieval database is built from a pre-recorded video taken along a planned walking path. Simple image retrieval using local image features does not work well when the images were taken in different conditions of time and weather. In this paper, we especially explore robustness for the change of sunlight condition due to its recording time. We propose a new method of robust location estimation based on our database built by combining more than one video that is taken under different conditions. Experiment results showed that the accuracy of the location estimation using the proposed database is better than the one from a conventional database built by using a single video.