Environmental Factors in Indoor Navigation Based on Real-World Trajectories of Blind Users
Hernisa Kacorri, Eshed Ohn-Bar, Kris M. Kitani, Chieko Asakawa · 2018 · Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3173574.3173630
Summary
This paper addresses a methodological blind spot in blind-navigation research: studies typically report localization accuracy and completion time, but say almost nothing about how the physical environment itself shapes where users go wrong. The authors argue that the same instruction ('walk straight') behaves completely differently depending on whether the user is in a narrow corridor with echoing walls or a wide open atrium, yet environmental context is rarely quantified. They propose two new, comparable measures of path-following quality: trajectory variability (how closely a user adheres to the planned route, computed via a path-normalized chi-squared distance between movement-orientation histograms) and deviation (human-annotated onsets of wandering off-path). Nine blind participants (8 white-cane users, 1 guide-dog user) navigated three multi-floor routes totaling 407 meters through a 21,000 m² Tokyo shopping mall instrumented with 218 BLE beacons and the NavCog 3 turn-by-turn app. Ground-truth sub-meter trajectories were extracted from frame-by-frame analysis of 360° video — a dataset the authors emphasize is the first of its kind at this scale for blind users. They then fit a multiple-regression model predicting variability and deviation from six feature classes: walls (distances and orientations in four compass directions), tactile paving, scene elements (doors, stairs, elevators, obstacles), path geometry (curvature, spatial derivative), localization error, and instruction type. Code, features, and data are released publicly at envfactors.github.io.
Key findings
Environmental and route characteristics explain substantial variance in path-following behavior — the full regression model reached adjusted R² of 0.44 for variability and 0.36 for deviation. The four variables with the largest effects on variability were all environmental: proximity to stairs (StairsDist), standard deviation of pillar distances (PillarDistStd), orientation of nearby east walls, and obstacle instructions. The most surprising result: localization error — the dominant variable reported in prior work — was not statistically significant for variability (only for deviation). This implies that further refining beacon accuracy will not by itself fix veering: the environment's shape does. Pillar layouts and open spaces were specifically called out as challenges (paths can be bypassed from multiple directions, multiplying user-choice variability). Obstacle-warning instructions reduced variability, confirming that alerting users about surroundings tightens their trajectories; 'forward' and 'U-turn' instructions, conversely, were positively associated with deviation, likely because users poorly estimate traversed distance during long forward stretches. The variability and deviation metrics captured complementary aspects — pillar and tactile features mattered for variability but not deviation — reinforcing that both should be reported together. Cross-validation (leave-one-floor-out) yielded MSE of 0.106 (variability) and 0.268 (deviation), suggesting reasonable generalizability.
Relevance
For practitioners deploying or evaluating indoor blind-navigation systems, this paper is an argument against the dominant 'improve the localization' frame. Once beacon accuracy is in the ~2-meter range that modern BLE systems already achieve, further gains come from understanding and compensating for the building itself — open atria, pillar forests, staircase edges, and long featureless corridors. Route designers and system integrators should document floor-plan characteristics (wall geometry, pillar density, tactile paving coverage, proximity of static obstacles to the planned route) when reporting evaluation results; without these, cross-study comparison is effectively impossible. The released open-source pipeline (envfactors.github.io) lets researchers automatically extract these features from floor-plan images. Limitations include the single-site single-culture sample (one Tokyo mall, 9 participants), the focus on static scene elements (dynamic crowds and moving obstacles are out of scope), and dependence on human-annotated deviation onsets. Still, the paper provides concrete, reusable metrics that accessibility researchers can adopt today to make indoor-navigation studies comparable and to steer future system design toward environment-aware interaction.
Tags: indoor navigation · turn-by-turn navigation · blindness · visual impairment · assistive technology · orientation and mobility · trajectory analysis · wayfinding · mobile accessibility · built environment · Bluetooth Low Energy · tactile paving · evaluation methods · research methodology
Standards referenced: WHO ICD-10 H54