Corridor-Walker: Mobile Indoor Walking Assistance for Blind People to Avoid Obstacles and Recognize Intersections
Masaki Kuribayashi, Seita Kayukawa, Jayakorn Vongkulbhisal, Chieko Asakawa, Daisuke Sato, Hironobu Takagi, Shigeo Morishima · 2022 · Proceedings of the ACM on Human-Computer Interaction (Proc. ACM Hum.-Comput. Interact., Vol. 6, MHCI, Article 179) · doi:10.1145/3546714
Summary
Walking an indoor corridor independently is deceptively hard for a blind traveller. Two problems compound: avoiding obstacles stacked against the wall (the same wall the traveller uses as a tactile guide) and recognising when an intersection has arrived and which directions it opens into. White canes detect low objects but miss overhanging furniture; guide dogs help but are scarce; previous research has offered either wearable obstacle-detection rigs (extra hardware, extra carrying cost) or turn-by-turn systems that depend on a pre-built static map of the building. Kuribayashi and colleagues propose Corridor-Walker, a system that runs on a single off-the-shelf iPhone 12 Pro with no precomputed map. The phone's LiDAR sensor continuously builds a 2D occupancy grid map of the corridor (with RANSAC floor-plane detection labelling cells as walkable, non-walkable, or unknown); an A* path planner generates an obstacle-avoiding route that keeps the user away from walls; a YOLOv3 detector trained on 9,940 grid-map images classifies upcoming intersections into L-shaped, T-shaped, rotated T-shaped, and X-shaped topologies. Feedback is multi-modal: bone-conducting headphones deliver spatialised-audio veering correction and TTS detour instructions, while the phone vibrates to signal both imminent collisions (continuous) and detected intersections (0.1 s pulses). The authors then evaluate the system with 14 blind participants across three tasks — identifying intersection shape, avoiding obstacles along a 15 m straight corridor, and navigating long routes mixing both challenges.
Key findings
Corridor-Walker produced robust quantitative improvements over a cane-only baseline. For intersection identification, the system raised correct-shape answers from 71.4% to 92.9% (L-shaped), 21.4% to 92.9% (T-shaped), 28.6% to 100% (rotated T-shaped), and 0% to 50% (X-shaped), with Wilcoxon signed-rank tests significant at p<0.01 for the three non-L shapes. Wall and obstacle contact dropped sharply in every task: e.g., on an L-shaped turn, wall contacts fell from 3.86 to 0.14 (p=0.004); on Route 3-1 (three intersections, three obstacles) obstacle contacts fell from 3.07 to 1.28 (p=0.01) and wall contacts from 12.21 to 1.07 (p=0.003). Intersections were detected on average 2.47 m before the user reached them, with precision and recall above 0.9 within 2.5 m. Mean SUS score was 80.5 (SD 7.41), comfortably in the 'good' range. The trade-off: system-aided walks took roughly 50–70% longer because participants slowed to process audio and re-orient. Qualitatively, 12 of 14 participants praised the obstacle-avoidance function ('I was able to avoid an obstacle without even knowing it was there'), and 9 of 14 said the intersection detector would help them build a mental map of unfamiliar buildings. Participants with strong echolocation skills (P03, P13) found the orientation-correction and intersection functions redundant for them, underlining the need for adjustable interfaces.
Relevance
For practitioners and researchers in indoor wayfinding, this paper is a strong argument that LiDAR-equipped consumer smartphones have reached a threshold where real-time, map-free indoor navigation assistance is practical — no wearable rig, no building-scale BLE-beacon retrofit, no tactile map distribution. That lowers the cost of deployment dramatically for hospitals, office buildings, apartments, and transit spaces where blind travellers need to cope with unfamiliar but locally known corridor layouts. The division of feedback (spatialised audio for orientation, TTS for deliberate instructions, vibration for urgency and intersection presence) is a practical template for other multimodal assistive systems. Important limitations: the system handles only perpendicular intersections with fixed corridor widths, its 5 m LiDAR range fails in open lobbies, holding the phone parallel to one's body is fatiguing (11 of 14 participants flagged this), and the O&M training required to use it well has not yet been designed. The finding that high-echolocation users may not benefit from every feature is a useful caution against one-size-fits-all assistive navigation.
Tags: indoor navigation · blind navigation · obstacle avoidance · intersection detection · LiDAR · orientation and mobility · visual impairment · smartphone assistive technology · wayfinding · spatialized audio · vibration feedback