PathFinder: Designing a Map-less Navigation System for Blind People in Unfamiliar Buildings
Masaki Kuribayashi, Tatsuya Ishihara, Daisuke Sato, Jayakorn Vongkulbhisal, Karnik Ram, Seita Kayukawa, Hironobu Takagi, Shigeo Morishima, Chieko Asakawa · 2023 · Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3544548.3580687
Summary
Map-based indoor navigation systems for blind travellers (CaBot, NavCog, BLE-beacon apps) work well when a prebuilt map exists, but building and maintaining those maps is labour-intensive and has only been done for a tiny fraction of the world's buildings. Engel et al. reported that 59.4% of blind people who travel to unfamiliar buildings ask a sighted person to accompany them — the most common workaround. Kuribayashi and colleagues ask: what would a navigation system that assumes no map look like, and what environmental information does a blind traveller actually need to reach a destination on their own? The authors begin with a scenario-based participatory design: five blind participants are walked along two unfamiliar office routes (46 m and 166 m) by an experimenter who acts as the robot, relaying route descriptions originally gathered from ten sighted passersby. Participants rate which indoor features are useful. Intersections, directional signs, and textual signs emerge as the top three. The team then builds PathFinder on a CaBot-derived suitcase robot fitted with a 360° LiDAR (Cartographer SLAM), an iPhone 12 Pro on an elevated stabiliser for higher-resolution sign capture, an NVIDIA RTX 3080 PC, and a four-button handle interface. The system combines Yang et al.'s convex-hull-and-skeletonisation intersection detection with a sign-recognition pipeline (iOS Vision OCR for real-time detection; EasyOCR plus YOLOv5 arrow detection and Laplacian-based rule-based grouping on demand) to read back directional signs (e.g., 'Right, Corridor 4200') and textual signs (e.g., room numbers). A design iteration with the same five participants simplified the feedback, replaced clock-position intersection cues with 'left/right/forward/backward', added a 'Take-me-back' button, and merged stop+sign-recognition onto one button. A main study with seven new blind participants (P06–P12) compared PathFinder against a prebuilt-map 'topline' system on the same two routes.
Key findings
All seven participants reached every destination using PathFinder. Confidence and cognitive-load ratings were significantly higher than their regular aid (p<.05 for Q1 and Q2) — and significantly lower than the topline (prebuilt-map) system (p<.05), placing PathFinder as an 'in-between' tool: better than cane-only travel, not as seamless as a fully mapped building. Task-completion times reflected the trade-off: mean normalised 317.4 s (R1) and 607.9 s (R2) with PathFinder vs 63.2 s and 239.6 s with the topline — PathFinder asks the user to stop at every intersection and decide, whereas the mapped system autoroutes. Mean SUS was 85.25 (grade A). Intersection detection produced 61 correct, 9 partial, 5 failed, and 33 false-positive detections across the study, with most false positives caused by a glass bridge (transparent to LiDAR) and by crowding near the R2 elevators. Sign recognition was initiated 62 times with 27 correct-and-relevant, 12 correct-but-irrelevant, 13 null, and 10 wrong results. The 'Take-me-back' function was unanimously praised; the guide-dog user P12 rated the topline system highest and favoured his dog for known routes but recognised PathFinder's unique value for unfamiliar ones. A usability issue to flag: on directional signs (e.g., 'Left, Corridor 4200') two participants pressed the left-turn button immediately, not realising the sign described a turn to be made later at the intersection, leading to accidental backward turns.
Relevance
This paper reframes indoor accessibility from 'map every building' to 'build a robot that reads the building'. That reframing matters because the per-building mapping cost is exactly what has stopped assistive navigation from scaling beyond a handful of museums, campuses, and transit hubs. The participatory finding — that intersections, directional signs, and textual signs together cover most of the environmental information blind travellers need — is also a useful simplification for future system designers: you do not need to model every obstacle and POI; you need to reliably locate decision points and the signage that describes them. Practitioners should pay attention to three weaknesses. First, transparent surfaces (glass bridges, glass doors, floor-to-ceiling windows) remain a systemic LiDAR failure mode and require RGB fusion. Second, the interface of 'read a directional sign aloud, then let the user act' proved ambiguous — participants conflated 'the sign says left' with 'turn left now'. Third, the system requires a 40-pound wheeled platform, a 2.6-hour battery, and currently ignores stairs and outdoor transitions. The authors explicitly invite exploration of smartphone-only and wearable forms for the same map-less approach, which is a natural research direction.
Tags: map-less navigation · blind navigation · indoor navigation · intersection detection · sign recognition · visual impairment · orientation and mobility · participatory design · shared control · assistive robotics · LiDAR · OCR