Introduction
Researchers have introduced SurDis, a specialized dataset designed to help wearable AI detect "surface discontinuities"—like steps, drop-offs, and open drains—that pose significant fall risks to blind and low-vision (BLV) individuals. By providing over 17,000 annotated depth maps captured from a pedestrian's perspective in urban Malaysia, the project fills a critical gap in assistive navigation technology.
The Problem: Navigating the "Drops"
Most computer vision research for autonomous navigation focuses on avoiding obstacles above the ground, such as furniture, pedestrians, or vehicles. However, for the BLV community, some of the most dangerous hazards are in the ground.
In many urban environments—particularly in low-to-middle-income countries—infrastructure is inconsistent. A sidewalk might suddenly end in a steep drop-off, an uncovered drainage ditch, or uneven stairs without handrails. Existing datasets (like those for self-driving cars) are often captured from vehicle heights or focus on well-regulated environments. There is a dire need for data that reflects the "first-person" reality of a pedestrian navigating challenging, unregulated urban terrain.
Methods: A Human-in-the-Loop Approach
The authors took a unique approach to define what actually constitutes a hazard for a BLV individual:
- Tactile Consultation: Since BLV volunteers couldn't "see" photos of hazards, the researchers created 3D-printed replicas of actual street scenes. Volunteers touched these models to identify which surface changes were most difficult to navigate with a white cane.
- Custom Wearable Hardware: They built a lightweight prototype consisting of a chest-mounted stereo camera (DUO MLX R2) and a single-board computer (Odroid-XU3). The camera was tilted 35 degrees downward to capture the immediate pathway.
- Data Processing & Annotation: To handle varying outdoor lighting, they applied Principal Component Analysis (PCA) whitening to normalize the images. They generated depth maps using a two-step block matching technique. Because depth maps are hard for humans to label accurately, they converted stereo pairs into red-cyan anaglyphs (3D images) to allow annotators to see the "drops" and "rises" clearly while drawing bounding boxes.
The SurDis Dataset
The paper introduces the SurDis (Surface Discontinuity) Dataset, which includes:
- Content: 200 sets of stereo image sequences resulting in 17,302 depth maps.
- Taxonomy: Five distinct classes of discontinuities: (1) down-steps, (2) up-steps, (3) uncovered drainage, (4) drop-offs without handrails, and (5) mixed gradients.
- Format: Annotations are provided in PASCAL VOC (XML) and COCO (JSON) formats.
- Comparison: Unlike datasets captured in highly regulated environments or via street-view cars, SurDis specifically targets hazardous, unregulated urban environments from a pedestrian's chest-height perspective.
Key Insights
- The "Cane Gap": Traditional aids like white canes often fail to detect steep drop-offs or wide open drains in time to prevent a fall, justifying the need for wearable electronic travel aids (ETAs).
- Perspective Matters: Data from cars or robots isn't transferable to BLV navigation. A chest-mounted camera captures the specific "looming" effect of a staircase or a curb as a human approaches it.
- Tactile Ground Truth: Using 3D printing to bridge the communication gap between sighted researchers and BLV users is a novel and highly effective way to ensure the dataset addresses real-world needs.
Key Results
The researchers established baseline performance using several popular object detection models:
- YOLOv4 emerged as the strongest baseline with a Mean Average Precision (mAP) of 0.427.
- Faster-RCNN followed with 0.396 mAP.
- SSD and ResNet-18 showed significantly lower performance (0.241 and 0.156 mAP, respectively).
These results suggest that while modern AI can identify these hazards, there is significant room for improvement, especially in distinguishing between "smooth" continuity and "hazardous" discontinuity in complex lighting.
Limitations and Future Work
- Geographic Scope: The data is currently limited to urban Malaysia. Surface conditions vary wildly across different countries and cultures.
- Algorithm Dependence: The quality of the depth maps depends on the specific stereo-matching algorithm used. The authors included the raw stereo bitmaps so future researchers can generate even more accurate depth data as algorithms improve.
- Physical Constraints: The data was captured by a person of a specific height; the "view" might change significantly for a shorter or taller user, or someone using a wheelchair.
- Real-time Feedback: While this paper focuses on the dataset, the ultimate goal—a real-time wearable system that provides haptic or audio feedback to a user—remains the next major engineering challenge.
Summary updated