Achieving Practical and Accurate Indoor Navigation for People with Visual Impairments
Dragan Ahmetovic, Masayuki Murata, Cole Gleason, Erin Brady, Hironobu Takagi, Kris Kitani, Chieko Asakawa · 2017 · Proceedings of the 14th International Web for All Conference (W4A) · doi:10.1145/3058555.3058560
Summary
This paper presents improvements to the NavCog indoor navigation system that achieve sub-meter localization accuracy for blind pedestrians while significantly reducing the infrastructure cost of deploying BLE (Bluetooth Low Energy) beacon networks. GPS provides outdoor localization but is ineffective indoors, where blind people face particular challenges navigating unfamiliar environments. NavCog uses a sparse network of BLE beacons installed in buildings, with smartphones detecting beacon signal strength (RSSI) fingerprints to estimate the user's position. The authors improve the original k-nearest neighbor (k-NN) regression approach by introducing two innovations: (1) modeling beacon RSSI probability distributions using kernel ridge regression, which enables accurate localization from fewer signal samples, and (2) integrating Pedestrian Dead Reckoning (PDR) using the smartphone's accelerometer and gyroscope within a particle filter framework. The PDR component detects steps via accelerometer peak detection and tracks orientation via gyroscope, providing continuous motion estimates between beacon readings. The navigation field is modeled as a graph of one-dimensional line segments (edges), reflecting how blind pedestrians actually navigate — following walls, curbs, and other linear reference features. This "Manhattan world assumption" dramatically reduces the number of beacons and signal samples needed compared to two-dimensional approaches.
Key findings
The improved system achieved 0.68m average localization accuracy with as few as 1 beacon every 6m and 0.5m sampling resolution — a major improvement over the k-NN baseline which averaged 1.2m error under the same conditions. Sub-meter accuracy was consistently reached with as few as 6 beacons (1 every 4m). The beacon selection analysis revealed that beacons at edge endpoints (path extremities) are most critical for initial localization, after which PDR efficiently maintains position tracking mid-path. User evaluation with 6 blind participants (ages 35-73) across two sessions six months apart showed statistically significant reduction in missed turns: from M=0.467 per segment (Session 1, k-NN) to M=0.267 (Session 2, improved method) on shared route segments (p=0.042). On non-overlapping segments, missed turns dropped from M=0.35 to M=0 (p=0.002). Qualitative feedback was strongly positive — participants noted the improved system better anticipated turns based on their walking pace, with one participant saying "I felt like I could almost do this on roller skates!" A guide dog user specifically noted that the improved timing allowed her to give her dog advance turning commands. Participants requested additional features: sound cues at turns, awareness of obstacles along hallway sides, preview mode, and re-routing capabilities.
Relevance
This research advances the practical deployment of indoor navigation for blind people by addressing the key barrier: installation cost. By achieving accurate localization with fewer beacons and less fingerprinting effort, the system makes it more feasible for building owners to instrument their environments. The workload analysis provides a concrete planning tool: for example, achieving 0.75m accuracy with 8 beacons requires choosing between 5 samples at 0.5m resolution (30s/m workload) or 25 samples at 4m resolution (8.8s/m workload). The user study provides compelling evidence that improved localization accuracy directly translates to better navigation outcomes — fewer missed turns and greater user confidence. The participant feedback highlights important design considerations: timing of turn instructions must account for individual walking speed, guide dog users need advance notice for dog commands, and users want environmental awareness beyond just directional guidance. The NavCog system is open source and has been deployed at university campuses, office buildings, and shopping malls, demonstrating real-world scalability. Limitations include the one-dimensional path model which cannot handle open spaces, and the need for more diverse user testing including younger participants.
Tags: indoor navigation · visual impairment · blindness · BLE beacons · Bluetooth Low Energy · localization · pedestrian dead reckoning · particle filter · NavCog · smartphone · wayfinding · assistive technology · user study · sensor networks · turn-by-turn navigation