Modeling Expertise in Assistive Navigation Interfaces for Blind People
Eshed Ohn-Bar, João Guerreiro, Dragan Ahmetovic, Kris M. Kitani, Chieko Asakawa · 2018 · Proceedings of the 23rd International Conference on Intelligent User Interfaces (IUI) · doi:10.1145/3172944.3173008
Summary
This short IUI paper asks a question most assistive-navigation research leaves unasked: what happens as a blind user becomes an expert on a route? Existing smartphone guidance apps deliver the same instruction set on a user's tenth trip down a corridor as on their first, ignoring a steady, predictable accumulation of route knowledge and interface familiarity. The authors study this evolution empirically with 8 blind white-cane users who repeatedly traversed two university-building routes (Route A: 500 ft, 8 turns, 13 POIs; Route B: 250 ft, 7 turns, 10 POIs) using NavCog 3 with BLE beacon localization. Each route was traversed four consecutive times with the app (T1–T4), interleaved with unassisted from-memory trials to probe retained route knowledge. User performance was measured along three dimensions: task completion time, route-description accuracy (FormElementError via Levenshtein distance, POIsMentioned, POIsCorrectLocation), and continuously tracked motion signals including walking speed and low-speed duration. The paper then proposes a personalized LSTM-based behavior model trained to classify whether a given trajectory segment represents novice (T1) or expert (T4) behavior, using position, velocity, angular velocity, and the current one-hot-encoded instruction type as features. The key methodological twist is personalization: the person ID is added as a one-hot feature and as an auxiliary classification loss in a multi-task formulation, on the grounds that expertise manifests differently across users.
Key findings
Route expertise evolved measurably across just four trials. Completion time dropped by an average of 45.2 s on Route A and 56.5 s on Route B from T1 to T4 (p<0.05 both routes), and walking speed rose from 0.54 m/s at T1 to 0.64 m/s at T4 (p<0.001). Participants spent less time stopped or walking below 0.2 m/s as trials progressed, reflecting reduced hesitation. Route structure recall improved significantly (FormElementError, p<0.005 both routes), though even after four exposures participants could only describe 26% of the route structure — rising to 69% after the final from-memory trial — suggesting guidance apps can suppress spontaneous learning unless trials force active memory use. POI mention rates stayed flat, but the fraction of POIs placed in the correct route segment rose significantly on Route B. The LSTM classifier with all cues reached 67.6% (Route A) and 77.7% (Route B) accuracy at distinguishing novice vs expert trajectories. Critically, removing the person-ID personalization dropped accuracy to 59.4% / 61.3% — a 17-point improvement from personalization alone — confirming that expertise is highly individual. The model performed best on POI/landmark instructions for Route A and turn instructions for Route B, showing that expertise shows up most clearly at interaction events, not during long forward walks.
Relevance
For designers of blind-navigation apps, the practical implication is clear: the interface should know whether it is talking to a novice or someone on their tenth walk through this corridor, and it should modulate verbosity, landmark detail, and approach-warning lead time accordingly. The personalized LSTM framework provides one path to do this automatically from ordinary sensor streams. The result that spontaneous route learning is suppressed when users lean on the app is particularly important for organizations deploying navigation aids: O&M training gains can be inadvertently reversed if users never exercise recall. Limitations are significant — 8 participants, only two short routes, no guide-dog or low-vision users, and expertise measured across four trials rather than weeks or months of real-world use. The classification accuracies (68–78%) also indicate that modeling user state from motion alone is hard, and more contextual cues (environment, cognitive load) will be needed in practice. Still, this is one of the few papers to put 'skill evolution' into the design vocabulary of assistive navigation, and it should inform how interfaces are built for repeated use of familiar spaces like the user's workplace, subway station, or frequently visited stores.
Tags: blind navigation · indoor navigation · turn-by-turn navigation · visual impairment · blindness · assistive technology · personalization · adaptive interface · user modeling · machine learning · deep learning · recurrent neural network · LSTM · route learning · mobile accessibility · orientation and mobility