Deep Learning Compensation of Rotation Errors During Navigation Assistance for People with Visual Impairments or Blindness
Dragan Ahmetovic, Sergio Mascetti, Cristian Bernareggi, João Guerreiro, Uran Oh, Chieko Asakawa · 2019 · ACM Transactions on Accessible Computing (TACCESS) · doi:10.1145/3349264
Summary
This paper addresses a critical but often overlooked problem in turn-by-turn navigation assistance for people with visual impairments or blindness (VIB): rotation errors at turning points. While much navigation research focuses on improving localization accuracy, this work demonstrates that even with perfect localization, users consistently over-rotate when following turn instructions, leading them onto incorrect and potentially dangerous paths. The system under study, NavCog, uses Bluetooth beacons for indoor localization and provides verbal turn instructions followed by an impulsive stop sound when the target angle is reached. The inherent delay between hearing the stop instruction and physically stopping causes systematic over-rotation. The authors analyze trajectory data from four diverse indoor environments — a shopping mall in Tokyo, the Andy Warhol Museum in Pittsburgh, Pittsburgh International Airport, and a hotel during a blindness conference — involving a combined 67 participants and nearly 500 rotations. They then propose a deep learning approach using a multilayer perceptron (MLP) neural network that predicts over-rotation in real time based on four features: current rotation angle, rotation duration, average rotation velocity, and maximum rotation velocity. The system anticipates the stop instruction, playing it before the target angle is reached so that the predicted over-rotation brings the user closer to the correct heading. The model was trained on a controlled dataset of 324 rotations from 18 blind participants performing turns on a swiveling chair.
Key findings
Rotation errors were found to be consistent and significant across all four environments, with average errors ranging from 14.6 degrees (airport) to 22.1 degrees (hotel). A key finding is that slight turns (22.5 to 60 degrees) produce significantly higher rotation errors than ample turns (60 to 120 degrees) across all environments, likely because smaller angles are harder to track vestibularly. In the mall dataset, 12 percent of turns resulted in incorrect navigation outcomes, with 42 percent of those requiring the user to stop navigation entirely. The deep learning compensation approach reduced average rotation errors from 30.1 degrees to 18.8 degrees in a user study with 11 blind participants — a 37.5 percent improvement that was statistically significant. Compensated rotations were also 8.8 percent faster than uncompensated ones, since the anticipatory approach does not add cognitive overhead the way continuous sonification alternatives do. Visual impairment onset was the only participant characteristic consistently affecting rotation errors, with those blind since birth performing more accurately. Guide dog users showed lower rotation errors than white cane users in wide-corridor environments like airports, suggesting that environmental context interacts with mobility aids in complex ways.
Relevance
This research highlights an important gap in accessible navigation technology: even technically accurate systems can fail users at the interaction level. For practitioners building navigation assistive technologies, the paper provides a concrete, implementable technique for improving turn accuracy without changing the user interaction paradigm or adding cognitive load. The finding that rotation errors are universal across environments validates this as a fundamental challenge rather than a context-specific quirk. The work also demonstrates that deep learning can be effectively applied to small accessibility-specific datasets to solve practical problems. For organizations deploying indoor navigation systems in public venues like airports, museums, or shopping centers, the compensation technique could meaningfully improve safety and reliability. The research underscores that accessibility solutions must account for the full human interaction loop, not just the technology side of localization and pathfinding.
Tags: navigation assistance · visual impairment · blindness · deep learning · turn-by-turn navigation · indoor navigation · rotation errors · over-rotation