WeAllWalk: An Annotated Data Set of Inertial Sensor Time Series from Blind Walkers
German H. Flores, Roberto Manduchi · 2016 · Proceedings of the 18th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS '16) · doi:10.1145/2982142.2982179
Summary
This paper introduces WeAllWalk, an openly accessible and annotated data set of inertial sensor time series collected from blind individuals walking through complex indoor routes. The data set addresses a critical gap in assistive navigation research: while step counting and dead reckoning algorithms have been developed extensively for sighted walkers, blind individuals exhibit markedly different gait patterns that can confound these algorithms. Cane users swing their cane-holding arm left and right, producing additional upper body rotation; blind walkers often "scuttle" rather than walk steadily; and without visual feedback, they frequently veer off straight paths and must correct course upon detecting a wall or obstacle. Eight blind participants (six long cane users and two guide dog users, aged 26-69) and five sighted controls walked six different paths across two university buildings. Routes ranged from 75 to 300 meters, included turns at 45, 90, and 180 degrees, required opening doors (push bars and pull handles), and traversed varied floor surfaces (carpet, linoleum, concrete). Each participant carried two iPhone 6s in self-selected pocket locations, recording accelerometer, gyroscope, and magnetometer data at 25 Hz. Ground truth heel strike times were captured by MetaWear-CPRO sensors clipped to participants' shoes. All time series were carefully annotated with intervals for straight segments, turns, and specific events such as obstacle collisions, being caught in wall openings, or stopping momentarily. The data set is released under Creative Commons CC-BY-4.0.
Key findings
The data set captures realistic blind walking behavior including involuntary veering, obstacle collisions, getting caught in wall openings, sporadic stops, and path corrections — events that are typical of blind ambulation but absent from existing gait data sets based on sighted walkers on straight paths. Detailed participant profiles reveal the diversity of blind mobility: cane techniques varied from synchronized sliding to non-synchronized swinging to echolocation-based tapping; guide dog users maintained straighter paths but experienced different challenges (dogs being distracted or making navigation errors). Step counting using the AMPD algorithm on smartphone accelerometer data showed higher error rates for blind participants compared to sighted ones, with both undercount and overcount events more prevalent, particularly for cane users whose arm movements create confounding acceleration signals. Turn detection using azimuth data was complicated by the less steady heading direction of blind walkers — heading variation could be as large as 20 degrees for cane users compared to relatively stable headings for sighted walkers, risking false positive turn detections. The data set includes over 30 hours of annotated sensor data across all participants and paths, with each recording accompanied by synchronized GoPro video for verification.
Relevance
WeAllWalk fills a significant gap in accessible navigation research by providing the first openly available inertial sensor data set specifically from blind walkers navigating realistic, complex indoor environments. For researchers developing smartphone-based indoor navigation systems for blind users, this data set enables algorithm testing against real-world blind walking patterns rather than sighted walking proxies that may not generalize. The rich annotations — marking not just steps and turns but also obstacle encounters, door interactions, and veering events — make the data useful for studying the full range of blind mobility challenges. For accessibility practitioners, the paper highlights important design considerations: many blind individuals prefer receiving distance information in steps rather than feet; step counting algorithms calibrated for sighted gait may significantly miscount for blind users; and the choice of mobility aid (cane vs. guide dog) substantially affects gait patterns and sensor data. The open CC-BY-4.0 licensing ensures broad reusability, and the detailed participant descriptions provide context for understanding individual variation in blind mobility.
Tags: blindness · indoor navigation · wayfinding · inertial sensing · data collection · orientation and mobility · long cane · guide dog · dead reckoning · open source