WeAllWalk: An Annotated Dataset of Inertial Sensor Time Series from Blind Walkers
Germán H. Flores, Roberto Manduchi · 2018 · ACM Transactions on Accessible Computing (TACCESS) · doi:10.1145/3161711
Summary
This paper introduces WeAllWalk, an annotated dataset of inertial sensor time series collected from blind and sighted walkers navigating complex indoor routes. Ten blind volunteers (seven using a long cane, one using a guide dog, and two alternating between both) and five sighted volunteers walked through fairly long indoor routes in two university buildings, encountering obstacles to avoid and doors to open. Inertial data were recorded by two iPhone 6s carried in participants's pockets, while ground-truth heel strike times were measured by small IMU units clipped to shoes. The dataset contains approximately 7 miles of walking data with around 20,000 annotated steps, carefully subdivided into straight paths and turns, with special events (bumping into obstacles, door openings) individually marked. The paper's primary contribution beyond the dataset itself is an in-depth comparative analysis of step counting and turn detection algorithms across the three walker communities: blind cane users, blind guide dog users, and sighted walkers. Six step counting algorithms (UPTIME, AMPD, WPD, ZC-acc, ZC-gyro, HMM-acc) and two HMM-based turn detection algorithms were evaluated using stratified leave-one-out cross-validation. The study was motivated by the fact that pedestrian dead reckoning — estimating position by counting steps and detecting turns from a known starting point — is a fundamental localization technique for indoor navigation systems, yet existing algorithms were developed and tested exclusively on sighted walkers.
Key findings
The analysis revealed significant performance differences across walker communities. For step counting, blind cane users had significantly higher error rates (approximately 50% higher) than sighted walkers, while blind guide dog users fell between the two groups. The best-performing algorithm for blind cane users was ZC-acc (zero-crossing on accelerometer data) with 7.8% mean error, compared to 20.4% for the worst performer (WPD). For sighted walkers, ZC-gyro performed best at 9.1%. Importantly, the same algorithm did not perform best across all communities — algorithm choice interacted significantly with walker community. For turn detection, HMM-Turn 2 outperformed HMM-Turn 1 (42.6% vs. 54.8% mean TD-Error), producing more balanced undercount and overcount rates. Turn detection errors for blind participants were approximately 10% higher than for sighted participants, though this difference did not reach statistical significance (p=0.09). Phone placement location (front pocket, back pocket, jacket, chest level) did not significantly affect either step counting or turn detection accuracy. The dataset annotations include not just steps and turns but also obstacle encounters and door interactions — events that produce distinctive inertial signatures that could enable automatic detection of navigation challenges.
Relevance
WeAllWalk addresses a critical gap in navigation research: the assumption that algorithms developed for sighted walkers will work equally well for blind users. The finding that step counting error rates differ significantly across communities — and that the best algorithm varies by community — has direct implications for any navigation system using pedestrian dead reckoning for blind users. Developers cannot simply apply standard pedometer algorithms and expect equivalent performance. The dataset itself is a valuable public resource for the accessible navigation research community, enabling algorithm development and benchmarking on representative data without requiring each research group to recruit blind participants. The distinction between cane and guide dog users adds important nuance: these mobility aids produce different gait patterns that affect sensor readings, suggesting that navigation systems should adapt their algorithms based on the user's mobility aid. For the broader AI fairness discussion, WeAllWalk exemplifies the benchmark dataset approach advocated by Guo et al. — creating representative evaluation data that includes people with disabilities to reveal performance disparities that aggregate metrics would hide.
Tags: indoor navigation · blindness · inertial sensing · step counting · turn detection · pedestrian dead reckoning · dataset · white cane · guide dog · gait analysis