Scaling Pedestrian Crossing Analysis to 100 U.S. Cities via AI-based Segmentation of Satellite Imagery

This paper presents a scalable AI-driven method using satellite imagery and the Segment Anything Model to automatically measure pedestrian crossing distances across America's 100 largest cities, revealing that older cities tend to have wider, more car-centric streets with median crossing distances ranging from 32 to 78 feet.

The Gist

Imagine you are trying to understand how easy it is to walk across a street in a city. The length of that walk—the distance from one sidewalk to the other—is a huge factor in whether people feel safe crossing or if they get hit by cars. But measuring this distance for every single intersection in a city is like trying to count every grain of sand on a beach; it's too big, too messy, and takes way too long for humans to do by hand.

This paper describes a clever way to use a "robot eye" (Artificial Intelligence) to measure these street crossings for the 100 biggest cities in the U.S. all at once. Here is how they did it, broken down into simple steps:

1. The Problem: Too Many Streets to Measure

For years, researchers have known that longer crossings are more dangerous. But we didn't have a map of crossing distances for the whole country. Previous attempts were like trying to paint a mural by hand: accurate, but incredibly slow and labor-intensive. They also mostly looked for painted "zebra stripes" on the road, missing the many crossings that have no paint at all.

2. The Solution: A Digital "Cut and Paste" Job

The researchers built a three-step assembly line to automate the process:

• Step 1: Taking the Photos (The Snapshot)
  They used a computer program to grab satellite photos of about 3 million street intersections across the 100 largest U.S. cities. Think of this as taking a bird's-eye snapshot of every crossroads in America.

• Step 2: Teaching the Robot (The Art Class)
  They needed the computer to know the difference between a road (where cars go) and a sidewalk (where people walk). To teach this, they showed the AI (called Meta's "Segment Anything Model") a small batch of photos where humans had manually colored in the sidewalks and buildings.

  • The Analogy: Imagine showing a child a picture of a cookie and a picture of a plate, coloring the plate blue and the cookie brown. Once the child learns the pattern, you can hand them a new picture, and they can instantly color the plate blue without you telling them again.
  • They taught the AI to spot "non-drivable" areas (sidewalks, parks, buildings) and ignore the drivable roads.

• Step 3: The "Grow-Cut" Magic (The Scissors)
  This is the most creative part. The researchers took a digital map (OpenStreetMap) that had rough lines indicating where crossings might be.

  • The Analogy: Imagine you have a piece of string laid across a table, but the string is too long and hangs off the edges. You have a pair of magic scissors that only cut the string when it touches a specific colored zone (the sidewalk).
  • The computer took the rough crossing lines from the map and "grew" them slightly. Then, it used the AI's "colored zones" (the sidewalks) as a guide to "cut" the lines exactly where the sidewalk begins. This gave them the precise distance from one side of the street to the other.

3. The Results: A National Map of Walking Distances

By running this process, they successfully measured nearly 800,000 crossings in about an hour per city.

• How Accurate is it?
  They tested it in San Francisco against data humans had verified by hand. The AI was 93% accurate. On average, the AI was off by only about 2 feet and 3 inches (less than a meter). That's like guessing the length of a car and being off by the length of a single step.

• What Did They Find?

  • Old vs. New Cities: Older American cities (founded before 1800) generally have shorter crossings. Newer cities (founded later) have much longer crossings. This suggests that as America grew, it started building wider streets designed for cars, making it harder for pedestrians.
  • Region Matters: Cities in the Northeast and Midwest tend to have shorter crossings (around 30 feet), while cities in the South and West have much longer ones (up to 78 feet).
  • The Pattern: In almost every city, most crossings are short (neighborhood streets), but there are "corridors" of very long crossings (big highways) that stand out.

4. Why This Matters

This study gives city planners a "superpower." Instead of guessing or spending years measuring streets, they now have a map showing exactly where crossings are too long. This helps them decide where to build safety islands or shorten sidewalks to make walking safer, especially for the elderly, parents with strollers, or anyone with mobility issues.

5. The Limitations (The "Gotchas")

The authors are honest about where their method isn't perfect:

• Tree Trouble: If a street is covered in thick tree leaves, the satellite camera can't see the sidewalk, so the AI might get confused.
• Map Gaps: The system relies on OpenStreetMap to know where to look for a crossing. If a crossing isn't on that map, the AI won't measure it.
• Missing City: They had to swap out Anchorage, Alaska, for a city in Texas because the satellite maps for Alaska weren't available in the format they needed.

In short, this paper shows how we can use a combination of satellite photos, a smart AI, and a digital map to instantly measure how "walkable" our cities are, revealing that newer American cities are built wider for cars, while older ones are tighter for people.

Technical Summary

Technical Summary: Scaling Pedestrian Crossing Analysis to 100 U.S. Cities via AI-based Segmentation of Satellite Imagery

Problem Statement
Accurate, granular data on street dimensions, specifically pedestrian crossing distances, is essential for evaluating urban design impacts on travel behavior and safety. However, existing datasets are often missing, outdated, incomplete, or not machine-readable. While previous computer-vision studies have successfully identified marked crosswalks (e.g., "zebra" stripes) in limited cities, they fail to capture unmarked crossings, which constitute a significant portion of intersections in American cities (e.g., 42% in San Francisco). Furthermore, previous manual or semi-automated methods for measuring both marked and unmarked crossings have been labor-intensive and lacked scalability to a national level. This study addresses the need for a scalable, accurate, and reproducible method to measure crossing distances across the 100 largest U.S. cities using only freely available data.

Methodology
The study employs a three-step automated pipeline combining OpenStreetMap (OSM) data with Meta's Segment Anything Model (SAM) to process approximately three million satellite imagery tiles (8.3 terabytes).

1. Intersection Image Segmentation:

   • Data Retrieval: Using OSMnx, coordinates for all street intersections and mid-block crossings in the 100 largest U.S. cities were generated. Satellite imagery tiles (25-meter radius, 1629x1629 pixels) were retrieved via the Google Maps API.
   • Model Training: Meta's SAM (a zero-shot model) was fine-tuned to differentiate drivable surfaces (roads) from non-drivable surfaces (sidewalks, buildings, parks, refuge islands).
   • Training Strategy: To optimize training, 193 high-entropy intersection images (selected for visual complexity) from 14 diverse cities were manually annotated. The dataset was augmented via flipping, rotation, blurring, and noise addition, expanding the training set to 5,790 images.
   • Post-Processing: To improve accuracy, two corrective steps were applied:
     • False Positive Reduction: Polygons overlapping with buffered OSM street centerlines (1.2m for untagged roads, 2m for primary roads) were removed.
     • False Negative Correction: Missing non-drivable polygons were supplemented by merging building footprints extracted from OSM.
   • Output: The model generated pixel-level masks of non-drivable surfaces, converted into polygonal geometries mapped to real-world coordinates.

2. Crossing Edge Extraction:

   • Pedestrian crossing edges were extracted from OSM.
   • Single-node crossings (common in OSM) were converted into edges by matching nodes to the nearest "highway" edges.
   • Duplicates were merged based on spatial proximity (within 5m) and heading alignment (within 5 degrees).

3. Grow-Cut Algorithm:

   • Extracted crossing edges were overlaid onto the segmented satellite imagery.
   • A grow-cut algorithm expanded each edge by 5% from its endpoints while maintaining orientation.
   • The algorithm "cut" the expanded edges where they intersected with the non-drivable surface polygons, yielding the final crossing distance.
   • Filtering: Final outputs were filtered to remove duplicates (within 2m buffers) and outliers (crossings <6 ft or >160 ft).

Key Contributions

• Scalability: The method successfully processed 808,377 pedestrian crossings across the 100 largest U.S. cities, with a processing time of roughly one hour per city.
• Inclusivity of Crossing Types: Unlike prior studies relying on visual markings, this approach captures both marked and unmarked crossings by leveraging the geometric relationship between road edges and non-drivable surfaces.
• Reproducibility: The entire workflow utilizes freely available data (OSM, Google Maps) and open-source models (SAM), with code made publicly available on GitHub.
• National Dataset: The study produces the first nationwide dataset of pedestrian crossing distances, enabling inter-city and regional comparisons previously impossible at this scale.

Results

• Accuracy: Validated against a manually verified dataset for San Francisco, the automated method achieved 93% accuracy with a median absolute error of 2 feet 3 inches (0.69 meters). The automated process covered 72% of the crossings present in the manual dataset, limited primarily by the availability of crossing edges in OSM.
• Spatial Patterns:
  • Regional Variance: Median crossing distances range from 32 feet (9.8m) to 78 feet (23.8m). Shorter crossings cluster in the Northeast and Midwest (e.g., Toledo, OH at ~32 ft), while longer crossings are prevalent in the Southwest and West Coast (e.g., North Las Vegas at ~78 ft).
  • Urban Form: Northeast and Midwest cities exhibit a peak distribution near 30 feet, likely due to rectilinear grids. Southern and Western cities show a peak near 40 feet, reflecting more sprawling layouts and wider arterial corridors.
• Historical Correlation: A positive relationship was identified between a city's year of incorporation and median crossing distance. Cities incorporated before 1800 generally have shorter crossings (~40 ft), while those incorporated later show a progressive trend toward longer, more car-centric street designs.

Significance and Claims
The paper claims that this methodology unlocks new insights into pedestrian networks at a national scale, demonstrating that crossing distance follows a universal "left-shifted normal distribution" across U.S. cities but is significantly modulated by region and urban history. The authors assert that these datasets empower planners and safety advocates to:

1. Identify specific neighborhoods and corridors for targeted traffic-calming investments.
2. Combine crossing data with other local attributes (speed limits, crash data, demographics) to inform street redesigns.
3. Test whether associations between crossing distance and pedestrian safety, previously observed in smaller studies, hold true at the national scale.

The study concludes that the increasing emphasis on wider, car-centric streets in newer American cities is quantitatively evident through longer median crossing distances. The authors note limitations, including reduced accuracy in areas with heavy tree canopy (which obscures sidewalks) and reliance on the completeness of OSM data, but emphasize the method's potential for adaptation to international contexts and other urban features like bicycle infrastructure.