Vision-based geolocation methods to assist outdoor navigation of persons with visual impairment

Busatto Figueiredo, Renato (2024) Vision-based geolocation methods to assist outdoor navigation of persons with visual impairment. Doctoral thesis, University of East Anglia.

Preview

PDF Download (52MB) | Preview

Abstract

Persons with visual impairment are affected daily by the lack of accessibility. In this thesis, we address the research question: Can we use computer vision techniques to improve the accuracy of geolocation estimation to potentially assist persons with visual impairment navigating outdoors? We analyse the requisites to create an outdoor visual navigation system and highlight the main problems involved. The main challenge identified is estimating an accurate user position. To tackle this problem, we detail the construction of the HLoc+SV, a vision-based geolocation method inspired by a version of the hierarchical localisation framework that exploits information from a set of geotagged street-view images. In a dataset of 58 pictures and 80 geotagged street-view reference images, HLoc+SV had a mean absolute geolocation error of 0.77m (SD 0.41), while a smartphone GNSS receptor had a 12.09m (SD 8.67) error. Nonetheless, the HLoc+SV is a potential solution suitable only for the scenario when the GNSS service is available and relatively accurate. When the GNSS service is unreliable or unavailable, we analyse a framework to geolocate an image using a GMCP-based image retrieval method combined with the Self Quotient Image (SQI) illuminance normalisation. We found out there is a degradation of 4% on results compared to the original method when a geometric transformation is estimated by combining images with and without SQI. We also propose a method to isolate and measure the impact of changes in illuminants using a graph-morphological algorithm known as Sieve. We investigate the impact of using SQI on local features by segmenting images by levels of detail. We found that flat areas produced by the sieve have a positive effect on the detection of MSER blobs. MSER combined with SQI and sieve proved to be robust in matching street view images, increasing the matching score by 90% in specific scenarios compared to SIFT features extracted from original images.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Science > School of Computing Sciences
Depositing User:	Chris White
Date Deposited:	25 Jun 2024 08:55
Last Modified:	25 Jun 2024 08:55
URI:	https://ueaeprints.uea.ac.uk/id/eprint/95663
DOI:

Downloads

Actions (login required)

View Item