Closing the Gap: Designing for the Last-Few-Meters Wayfinding Problem for People with Visual Impairments
Manaswi Saha, Alexander J. Fiannaca, Melanie Kneisel, Edward Cutrell, Meredith Ringel Morris · 2019 · Proceedings of the 21st International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS 2019) · doi:10.1145/3308561.3353776
Summary
This paper comprehensively investigates the last-few-meters wayfinding problem — the gap between where GPS navigation ends (within approximately 5 meters of a destination) and where a person with visual impairment actually needs to arrive (the specific door, entrance, or storefront). GPS can bring users to the vicinity of their destination but not to the precise location, creating confusion and dependence on sighted assistance. The research consists of two studies. The first is a formative online survey (N=22, ages 18-50+, 15 totally blind, 7 with some residual vision) investigating challenges, current resolution techniques, and information needs. The second is a design probe study (N=13, ages 24-55) using Landmark AI, a camera-based iOS app modeled on Microsoft Seeing AI that provides three channels of information: a Landmark channel (real-time object recognition of structural landmarks like doors, stairs, and elevators using on-device SqueezeNet), a Signage channel (OCR of captured images via Microsoft Azure Cognitive Services), and a Place channel (custom landmark capture and matching, implemented via Wizard of Oz). The app was designed to complement existing GPS navigation (Microsoft Soundscape) and users's O&M skills rather than replace them. Five categories of landmarks were identified from O&M literature: structural (doors, elevators), sound (fountains, bells), tactile (carpet changes, curb ramps), air (HVAC, fans), and smell (bakeries, perfume stores).
Key findings
The formative study revealed that finding the intended doorway was the hardest part of the last few meters (reported by 11/22 participants), caused by GPS inaccuracy, guide dog limitations, and missing or inaccessible signage. Medical centers were the most commonly cited problematic destination (6 participants). Participants resolved challenges through sighted assistance (17), O&M skills (11), trial and error (7), and technology (2). Tactile landmarks were most preferred (Mdn=5/5), followed by structural (Mdn=5) and sound (Mdn=4) landmarks. The design probe study at a large outdoor shopping center found that all 13 participants valued Landmark AI's information, citing faster mobility, increased independence, and reduced need to ask for help. The Landmark channel was most useful for real-time contextual awareness. Door detection was unanimously the most valued capability — participants emphasized differentiating doors from full-pane windows (which confuse residual vision, guide dogs, and cane users alike). The Place channel was the most liked (9 participants) for its ability to capture and share specific locations socially. The Signage channel was appreciated for reading signs too far away to see but had usability challenges around knowing when and where to point the camera. Key design implications include: systems must adapt to users' residual vision type (color perception, peripheral vs. central), mobility aid (cane users benefit from structural landmarks while guide dog users need tactile ones), confidence level, and situational context (noise, crowds, familiarity). Hands-free wearable form factors were requested by multiple participants since holding a phone conflicts with cane or guide dog use.
Relevance
This paper names and rigorously investigates a problem that every blind GPS user encounters but that had received no comprehensive research attention: the last few meters. The design space framework — organized around visual abilities, mobility aid, personality/preferences, and context — provides a practical blueprint for building adaptive navigation systems. The insight that different mobility aids create different landmark affordances (guide dogs for search tasks, canes for discovery tasks) is directly actionable for developers. The preference for precision over recall in object recognition (false positives are more harmful than missed detections) is a critical design principle for any AI-based accessibility tool. The social dimension of the Place channel — capturing and sharing specific meeting points — suggests navigation apps should support collaborative wayfinding, not just solo use. For accessibility practitioners, this work demonstrates that the last-few-meters problem requires a fundamentally different approach from long-distance navigation: it needs landmark recognition, environmental awareness, and egocentric spatial descriptions rather than turn-by-turn directions. The research also highlights that AI-based solutions should complement rather than replace existing O&M skills, empowering users rather than creating dependency.
Tags: blindness · wayfinding · navigation · GPS · computer vision · landmarks · orientation and mobility · last mile · object recognition