This survey paper, "From White Canes to Edge AI," provides a comprehensive look at the technological evolution of navigation aids for blind and low-vision (BLV) individuals. Spanning nearly two decades of research (2008–2026), the authors map out how we moved from simple mechanical tools to sophisticated, on-device artificial intelligence, while highlighting a sobering reality: despite massive algorithmic leaps, real-world adoption remains remarkably low.
TL;DR
The paper categorizes BLV assistive technology into four distinct "eras," tracing the shift from passive tools to cloud-dependent deep learning and, finally, to modern "Edge AI." It argues that the primary bottleneck to adoption isn't detection accuracy, but rather the failure to account for human cognitive load and a lack of user-centered co-design.
What Problem Does This Paper Address?
Independent mobility is a cornerstone of autonomy, yet for most of history, the only widely adopted solutions have been the white cane and the guide dog. While effective, these tools have "semantic deficits"—they can tell a user something is there, but not what it is (e.g., a trash can vs. a pedestrian).
Modern Electronic Travel Aids (ETAs) have attempted to fill this gap using sensors and AI. However, the authors identify a "60% prototype gap": most systems never advance beyond laboratory prototypes. The paper tackles the "why" behind this failure, pointing to three main issues:
- Infrastructure Dependence: Many systems rely on expensive beacons or stable cloud connectivity.
- Cognitive Overload: AI that constantly "shouts" alerts can drown out the natural acoustic cues (like traffic flow) that BLV users rely on.
- The Co-design Deficiency: Only 39% of reviewed projects involved BLV individuals during the initial design phase.
Methods
The authors conducted a systematic literature review of 108 peer-reviewed studies using the PRISMA framework.
Beyond just cataloging papers, they provide a deep technical analysis of:
- Detector Families: Comparing R-CNN, YOLO, SSD, and DETR through the lens of "edge-mobility verdicts"—essentially, can these models actually run on a smartphone in real-time?
- Model Optimization: A consolidated reference for techniques like pruning, quantization (e.g., converting 32-bit weights to 8-bit), and knowledge distillation to make heavy AI models "edge-native."
- Sensor Fusion: How systems combine data from cameras, ultrasonic sensors, IMUs, and GPS to reduce uncertainty.
The paper also positions itself against previous surveys, noting its unique focus on the technical "how-to" of edge deployment.
Datasets
The survey notes a significant problem with "benchmark fragmentation." While many researchers use general-purpose datasets like MS COCO or ADE20K, these often lack the specific hazards BLV users care about (like curbs or head-level obstructions).
Key datasets mentioned include:
- VizWiz: A "BLV-authentic" resource featuring low-quality, blurred, or poorly framed images taken by real users.
- NYU Depth v2: Used for training depth-estimation models.
- Custom Datasets: Many researchers build small, non-standardized sets for specific hazards, making it nearly impossible to compare the performance of different systems.
Key Insights and Ideas
- The "Simplicity Constraint": The white cane endures because it is reliable and doesn't block the user's hearing. Successful AI must respect this "auditory channel" and avoid overwhelming the user with notifications.
- The Cloud is a Hazard: Offloading AI to the cloud introduces latency (often 500ms+). At a normal walking pace, a user moves half a meter in that time—a delay that makes the system unsafe for real-time obstacle avoidance.
- Edge-Native is the Future: The "Era IV" of technology focuses on moving the intelligence onto the device itself (phones or smart glasses) to ensure privacy, reduce latency, and work in "internet-dead" zones like basements or subways.
- Integration vs. Fusion: Most systems merely integrate sensors (using them for separate tasks). True fusion (mathematically combining data to reduce error) is more robust but computationally harder.
Key Results
- YOLO Dominance: The YOLO (You Only Look Once) family is the most popular for assistive tech because its single-stage architecture is inherently faster than two-stage models like R-CNN.
- Optimization Payoff: Techniques like int8 quantization can more than double inference speed on mobile hardware with a negligible (<5%) drop in accuracy.
- The User Study Gap: In a cross-corpus comparison of 22 representative systems, the authors found that the majority were evaluated only on benchmark metrics. Only five reported testing with a significant number of confirmed BLV participants.
Limitations and Future Work
The authors outline several "open challenges" that the research community has yet to solve:
- Seamless Transitions: No current system handles the handoff between GPS-guided outdoor navigation and infrastructure-free indoor navigation gracefully.
- Negative Obstacles: Detecting "drop-offs" (like a downward staircase) remains a major safety gap, with some fused systems only reaching 67% accuracy.
- Privacy-Preserving Perception: Always-on wearable cameras raise massive ethical concerns for bystanders. The paper suggests "privacy frictions" like physical shutters or audible tones as potential mitigations.
- Multimodal Foundation Models: The next frontier is moving from "Object: Chair (90%)" alerts to conversational AI that can answer complex questions like "Where is the door handle?" or "Is this the right bus?"
Summary updated