Smartphone-Based Inertial Odometry for Blind Walkers.

Area of Science:

• Assistive Technology
• Human-Computer Interaction
• Robotics

Background:

• Smartphone inertial sensors offer potential for pedestrian tracking and navigation for blind individuals.
• Existing studies predominantly use data from sighted participants, potentially limiting applicability to blind walkers with distinct gait patterns.
• The WeAllWalk dataset is the sole published indoor inertial sensor dataset from blind walkers.

Purpose of the Study:

• To comparatively assess pedestrian tracking algorithms using inertial sensors for blind individuals.
• To evaluate algorithm performance in both map-available and map-unavailable indoor environments.
• To highlight the necessity of using data from blind walkers for developing robust assisted navigation systems.

Main Methods:

• Utilized the WeAllWalk dataset, comprising indoor inertial sensor data from blind walkers.
• Developed and evaluated a novel two-stage turn detector combined with an LSTM-based step counter for path reconstruction without a map.
• Compared the proposed method with RoNIN, a deep learning-based algorithm, and experimented with particle filtering and mean shift clustering when a map is available.

Main Results:

• The proposed two-stage turn detector and LSTM step counter demonstrated robust path reconstruction capabilities in map-unavailable scenarios.
• Particle filtering with mean shift clustering showed promise in map-available situations.
• Performance varied significantly, underscoring the unique challenges posed by the gait of blind walkers.

Conclusions:

• Training and testing inertial odometry systems with data specifically from blind walkers is crucial for successful assisted navigation.
• Current algorithms require adaptation and validation using datasets like WeAllWalk to meet the needs of blind pedestrians.
• Further research is needed to optimize algorithms for diverse mobility aid usage (e.g., long cane, guide dog).