United States Department of Transportation
i
Wearables to Command More Access and Inclusion in a Smarter Transportation System
-
2022-11-01
-
[PDF-4.78 MB]
Details:
-
Creators:
-
Corporate Creators:
-
Corporate Contributors:
-
Subject/TRT Terms:
-
Resource Type:
-
Geographical Coverage:
-
Corporate Publisher:
-
Abstract:Visual place recognition (VPR), technology often associated with navigation of autonomous vehicles, can be critical to meeting every day urban navigation needs of people with vision disabilities. This research addresses two major obstacles to implementing VPR at scale: 1) the need for side-view place recognition, crucial for identification of sidewalk features like storefronts; and 2) privacy concerns that result from capture of street-view images during the most relevant peak commute hours, and potential tension between obfuscation and inaccuracy that must be addressed before VPR database and query construction. Using an open-source dataset consisting of more than 200,000 images captured via camera-mounted taxis over a 2km by 2km area in Manhattan, New York, over the course of one year, researchers present benchmark results of the performance of popular VPR algorithms at both of these challenges. Results indicate that side-view recognition is significantly more challenging for current VPR methods, and that data anonymization has a negligible, or even marginally beneficial effect on performance. This research contributes to the larger body of research in the following ways: • Benchmarks VPR methods using a unique large-scale dataset of over 200,000 front-view and side-view images over a full year, capturing seasonal and other environmental variation • Analyzes the causes of the significant challenges of VPR approaches using side-view images • Using pixel removal as an anonymization technique and demonstrating that this anonymization has negligible impacts to VPR algorithm performance.
-
Format:
-
Funding:
-
Collection(s):
-
Main Document Checksum:
-
Download URL:
-
File Type:
