TLDR: RAMP NET is a two-stage AI pipeline that significantly improves curb ramp detection in streetscape images. Stage 1 automatically generates a large, high-quality dataset by translating government-provided curb ramp locations into pixel labels on Google Street View panoramas. Stage 2 then uses this dataset to train a deep learning model that achieves near human-level accuracy in detecting curb ramps, far surpassing previous methods, enabling scalable and low-cost urban accessibility audits.
Curb ramps are essential for urban accessibility, providing crucial pathways for individuals with mobility disabilities, parents with strollers, and travelers with luggage. Despite their importance, accurately identifying and mapping these ramps in streetscape images has been a persistent challenge. Traditional methods, such as manual “boots-on-the-ground” inspections, are often prohibitively expensive and time-consuming for municipalities. While computer vision techniques have been explored, their effectiveness has been limited by a scarcity of large-scale, high-quality datasets.
Addressing this critical gap, researchers have introduced RAMP NET, an innovative two-stage pipeline designed to significantly advance curb ramp detection. This system leverages open government metadata to automatically generate a vast, high-quality dataset, which then powers a state-of-the-art detection model. The project’s code and datasets are open source, encouraging further development and standardization in the field. For more details, you can refer to the full research paper.
Stage 1: Building a Comprehensive Dataset
The first stage of RAMP NET focuses on automatically creating a labeled dataset of curb ramps from existing government data. Many local governments possess metadata about curb ramp locations, typically as latitude and longitude coordinates. However, this data lacks corresponding images, making it unsuitable for direct use with computer vision models.
RAMP NET’s Stage 1 overcomes this by translating these geographical coordinates into pixel labels on Google Street View (GSV) panoramas. The process involves several steps: identifying relevant GSV panoramas within a 10-meter radius of known curb ramp locations, then selecting all curb ramps within a larger 35-meter radius of the panorama’s location as label candidates. The system also ensures that the curb ramp installation date precedes the panorama capture date to maintain accuracy. To reflect real-world scenarios, the dataset also includes “null images” – panoramas with no curb ramps – to prevent false positives.
A key innovation in this stage is the “auto-translation” method. For each potential curb ramp, the system extracts a directional image crop from the panorama. A modified ConvNeXt V2 deep learning model then analyzes this crop to pinpoint the exact pixel coordinates of the curb ramp. This model was initially pre-trained on Project Sidewalk data and further refined with a smaller, manually labeled dataset to improve performance.
This meticulous process has resulted in an unprecedented dataset comprising over 214,000 fully labeled panoramas and nearly 850,000 curb ramp labels. This scale and comprehensiveness far exceed previous efforts, providing a robust foundation for training advanced detection models. When evaluated against manually labeled panoramas, the Stage 1 generated dataset demonstrated impressive accuracy, achieving 94.0% precision and 92.5% recall.
Stage 2: State-of-the-Art Curb Ramp Detection
With the large-scale, high-quality dataset generated in Stage 1, the second stage of RAMP NET focuses on training a powerful curb ramp detection model. Crucially, unlike Stage 1 which relies on government metadata, the Stage 2 model operates solely on image data, meaning it can detect curb ramps in any city where Google Street View imagery is available, without needing pre-existing government location data for that specific area.
The model again utilizes a modified ConvNeXt V2 architecture, pre-trained on the ImageNet-1k dataset to enhance its learning capabilities. It is designed to perform heatmap regression, where it takes a full GSV panorama and outputs a heatmap indicating probable curb ramp points. Peaks in this heatmap represent detected curb ramps.
The model was trained on the extensive dataset created in Stage 1, incorporating data augmentation techniques like horizontal flipping to improve generalization. When rigorously tested against a manually labeled ground truth dataset, the RAMP NET detection model achieved an average precision (AP) of 0.924. This performance significantly surpasses previous state-of-the-art methods, which typically yielded AP values around 0.380, effectively bringing curb ramp detection to near human-level accuracy.
Also Read:
- StreetViewAI: A New Horizon for Accessible Street View Exploration
- Developing a Smart System for Identifying Door Types in Floor Plans
Impact and Future Directions
RAMP NET represents a significant leap forward in urban accessibility assessment. By providing a low-cost, scalable, and highly accurate method for detecting curb ramps, this research has the potential to transform how cities conduct accessibility audits. Instead of months-long manual inspections, entire cities could potentially be audited in a single day, allowing for more efficient planning and maintenance of accessible infrastructure. This also empowers accessibility advocates by providing tools to ensure compliance with regulations like the Americans with Disabilities Act.
While groundbreaking, the researchers acknowledge certain limitations and areas for future work. The current dataset is primarily derived from three U.S. cities, which may introduce regional biases. Future efforts could expand coverage to other regions with different curb ramp styles. Additionally, the system’s reliance on streetscape imagery means its accuracy and timeliness can be affected by the availability and recency of GSV data, especially in rural or rapidly changing areas.
Future research aims to extend this work beyond mere detection to include automatic quality assessment of curb ramps (e.g., steepness, presence of tactile warnings), which often requires richer spatial information like bounding boxes rather than just single points. Translating the detected pixel coordinates back into geographical coordinates for city planning workflows is another important direction. Furthermore, the pipeline could be adapted to detect other vital urban accessibility features, such as pedestrian signals, missing curb ramps, or path obstacles, to provide a more holistic understanding of urban accessibility conditions.
In conclusion, RAMP NET provides a robust foundation for fully automatic urban accessibility assessment, offering a novel technique for data generation, the largest and most comprehensive curb ramp detection dataset to date, and establishing new performance benchmarks for the field.
