Guiding Blind Pedestrians in Public Spaces by Understanding Walking Behavior of Nearby Pedestrians
Seita Kayukawa, Tatsuya Ishihara, Hironobu Takagi, Shigeo Morishima, Chieko Asakawa · 2020 · Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (IMWUT), Vol. 4, No. 3, Article 85 · doi:10.1145/3411825
Summary
Collisions with other pedestrians are one of the most common hazards for blind travellers in public space — 87.8% of blind people report colliding or nearly colliding with pedestrians, bicycles, or obstacles. White canes detect contact rather than approach, and guide dogs cannot always anticipate distracted sighted pedestrians glued to their smartphones. Prior work had proposed obstacle-avoidance systems, but almost all of them treat other people as static blockers to be routed around — the 'off-path' pattern. Kayukawa and colleagues argue this is a narrow view of how sighted pedestrians actually behave: sighted people blend two responses to approaching others, adjusting walking speed on the same line ('on-path') and detouring only when necessary ('off-path'). The authors design a suitcase-attached guiding system that does both. Two Intel RealSense D435 RGBD cameras provide a 135° × 42.5° field of view; a LiDAR sensor plus IMU feeds a ROS cartographer for SLAM-based localisation. A YOLOv3 pedestrian detector (trained on COCO), combined with a Kalman filter for trajectory tracking and the Hungarian algorithm for matching, predicts where each nearby pedestrian will be in four seconds. The system continuously emits low-urgency alerts when a predicted trajectory intersects the user's dynamic warning triangle and high-urgency alerts when someone enters a fixed 1.2 × 2.5 m emergency rectangle. Two feedback modalities were evaluated — an audio interface (beep patterns for on-path, TTS 'Right'/'Left'/'Go straight' for off-path) and a tactile interface (vibration motors plus a servo-driven 'directional lever' that physically rotates to point the correct direction). Fourteen blind participants tested both in a controlled environment and along a 180 m real-world route on an office ground floor.
Key findings
All 14 participants reached the goal without collision in the controlled environment and avoided 10+ potential collisions per trial in the real-world environment. Mean SUS scores were 76.1 (audio) and 75.0 (tactile), both rated in the 'B' range. For off-path navigation, the directional lever guided participants significantly faster than the speech-based audio interface (mean 53.6 s vs 59.5 s, Wilcoxon p=0.015). Participants showed significantly stronger preference for the tactile interface overall (Q6 vs Q5, p=0.013): 11 of 14 said they never use a headset while walking because they rely on ambient sounds (footsteps, echoes, engines) for safety. Audio alerts were rated clearer per alert ('beep sounds are distinguishable by pulse rate and pitch'), but the headset that delivers them disrupted the very auditory awareness participants use to stay safe. The tactile interface had its own weakness: in the real-world environment with rough tile flooring, 8 of 14 participants said the vibration alerts were hard to distinguish from floor-texture vibration travelling up the suitcase handle, significantly lowering the effectiveness rating for Q2 (on-path mode with tactile, p=0.003). Three cases of missed alerts occurred: two participants ignored vibration signals and one was misdirected by a suitcase-orientation mismatch. Nine of 14 participants said the suitcase form was too heavy or bulky for daily carry, but all 14 wanted the directional lever miniaturised onto a cane or hand device.
Relevance
This paper is a useful corrective for the common assumption in assistive navigation that 'avoid pedestrians' means 'route around stationary obstacles'. The on-path / off-path distinction is both a conceptual contribution — matching how sighted pedestrians actually navigate crowds — and a practical interface requirement, since pure rerouting systems force blind users off their familiar non-visual landmarks and risk disorientation. For practitioners designing collision-avoidance systems, three concrete takeaways stand out. First, headset-delivered audio alerts are often unacceptable to experienced blind travellers because they mask exactly the ambient sounds used for spatial awareness; tactile and shape-changing interfaces deserve more attention. Second, vibration feedback is sensitive to floor texture and device-body coupling, so vibration motors should probably be worn on the body rather than mounted on a rolling platform. Third, the 'directional lever' — a physical servo that rotates to indicate the correct bearing — was uniformly preferred and deserves serious study as a replacement for turn-by-turn TTS. Limitations include small sample size, a rolling-suitcase form factor most participants called too heavy, and the absence of a baseline comparison against cane or guide-dog-only navigation.
Tags: blind navigation · collision prediction · pedestrian avoidance · visual impairment · audio interface · tactile interface · indoor navigation · orientation and mobility · assistive robotics · suitcase robot · LiDAR