Deep Learning Compensation of Rotation Errors during Navigation Assistance for People with Visual Impairments or Blindness

Navigation assistive technologies are designed to support people with visual impairments during mobility. In particular, turn-by-turn navigation is commonly used to provide walk and turn instructions, without requiring any prior knowledge about the traversed environment. To ensure safe and reliable guidance, many research efforts focus on improving the localization accuracy of such instruments. However, even when the localization is accurate, imprecision in conveying guidance instructions to the user and in following the instructions can still lead to unrecoverable navigation errors. Even slight errors during rotations, amplified by the following frontal movement, can result in the user taking an incorrect and possibly dangerous path. In this article, we analyze trajectories of indoor travels in four different environments, showing that rotation errors are frequent in state-of-art navigation assistance for people with visual impairments. Such errors, caused by the delay between the instruction to stop rotating and when the user actually stops, result in overrotation. To compensate for over-rotation, we propose a technique to anticipate the stop instruction so that the user stops rotating closer to the target rotation. The technique predicts over-rotation using a deep learning model that takes into account the user's current rotation speed, duration, and angle; the model is trained with a dataset of rotations performed by blind individuals. By analyzing existing datasets, we show that our approach outperforms a naive baseline that predicts over-rotation with a fixed value. Experiments with 11 blind participants also show that the proposed compensation method results in lower rotation errors (18.8° on average) compared to the non-compensated approach adopted in state-of-the-art solutions (30.1°).