PASDA orthoimagery
Pennsylvania's open imagery program provides recent, high-resolution tiles. We use the most current vintage available for each study area so crosswalk paint is captured as it actually exists today.
Method · Data pipeline
From pixels to planning data.
Two computer-vision models, three output layers, one common thread: every detection is traceable back to a piece of public imagery a planner could open and verify in under a minute.
At a glance
The pipeline, in one frame.
Before we walk through inputs, models, and outputs separately. Here's the whole thing in a single picture: what gets detected on the street, what's measured from above, and what lands in your GIS as planner-ready summaries.
Inputs
Two public imagery sources, sampled deliberately.
Google Street View imagery
For each detected intersection, we sample several headings facing into the crossing. This gives the object detector a fair shot at every approach instead of biasing toward one direction.
The three pipelines
Two computer-vision models, one planning interface.
Each pipeline owns a single, defensible output. We resisted the urge to merge them into a single model; separate pipelines are easier to retrain, audit, and explain.
YOLO detection
Fine-tuned object detection extracts stop signs and traffic lights from intersection-facing imagery. Outputs are point features with confidence scores, exported per neighborhood as GeoJSON and CSV.
- Stop-sign F1: 0.77
- Traffic-light F1: 0.58
- Outputs include detection confidence and Street View metadata
U-Net segmentation
Crosswalk markings are segmented from 15 cm/pixel aerial imagery to produce polygon outputs at scale. The U-Net architecture handles per-pixel classification well, even on faded paint.
- Crosswalk F1: 0.94
- Precision 0.93 · Recall 0.95
- Deployable across multiple Philadelphia study areas
Approach-level measurement
Perpendicular transects estimate curb-to-curb crossing width to create a screening metric planners can use without commissioning a survey crew.
- Intersection-approach estimates in baseline CSV outputs
- Median width plus spread from sampled transects
- Designed for prioritization, not engineering survey replacement
Process
How a tile becomes a planning layer.
The diagram is intentionally human-scale. Anything more complicated would obscure the only question that matters in a planning meeting: where did this number come from?
Imagery → models → planner-ready layers. The crossing-width step (not shown) reuses U-Net polygons plus intersection geometry to compute curb-to-curb transects.
Results
Deployment scale, model fit, and screening value.
The numbers below are pilot-area results. They're meant to set realistic expectations before you trust an output in a memo, not to oversell.
Crosswalk segmentation
U-Net on PASDA tiles, evaluated against hand-labeled validation tiles.
0.94F1 score
0.93Precision
0.95Recall
Traffic-control detection
YOLO fine-tuned on Street View, evaluated by class.
0.77Stop-sign F1
0.58Traffic-light F1
3,745Total detections
Coverage
Pilot deployment across the project's eight Philadelphia neighborhoods.
9,950Street View images
3,610Crosswalk polygons
8Neighborhoods
What planners get
Outputs designed for GIS workflows, not just a demo.
Traffic-control points
Neighborhood-level GeoJSON for stop signs and traffic lights, each tagged with confidence + Street View link.
Intersection points
Detected intersections plus the Street View sampling points the model used, easy to inspect and validate.
Crossing widths
Approach-level measurements suitable for flagging unusually wide or high-variance crossings citywide.
Crosswalk polygons
U-Net polygon layers from aerial imagery, the foundation for future condition monitoring as imagery refreshes.
Limits and next steps
This is a screening system, not a final audit.
Current limits
- Pilot-area training data. Models were tuned on a subset of Philadelphia. Performance on architecturally distinct neighborhoods is untested.
- No citywide ground-truth audit yet. Validation was done on hand-labeled tiles, not field-walked corridors.
- Lower recall for traffic lights in complex scenes; backlight, foliage, and far-away hardware all degrade detection.
- Crossing widths are screening estimates, not engineering survey replacements.
Next steps
- Expand validation with field-walked corridors and crash-data joins.
- Connect outputs to accessibility datasets: curb ramps, ADA inventories, transit stops.
- Extend the workflow to additional cities and street types beyond Philadelphia.
- Add a feedback loop where planners can flag false positives, feeding the next training cycle.