Stop signs
0.77F1 score
3,745Total traffic-control detections
Stop signs are well-suited to YOLO: distinct shape, predictable color, usually unobstructed.
YOLO · Object detection
YOLO: spotting things in a single look.
We use YOLO to find stop signs and traffic lights in Street View imagery. Here's what it is, why it's fast, and the honest story of where it helps and where it stumbles.
Plain-language definition
What is YOLO?
YOLO is an object detector. Feed it an image and it returns a list: "stop sign, this box, 92% confident. Traffic light, that box, 64% confident."
The name stands for "You Only Look Once." Older detectors scanned an image multiple times: first proposing regions to consider, then deciding what was in each. YOLO does both jobs in a single forward pass. That's why it's fast enough for real-time video, and why we can run it across thousands of Street View frames without breaking the budget.
If U-Net is a careful colorist, YOLO is a quick spotter. It's not painting the object. It's putting a box around it and naming it before moving on.
Want to go deeper? Visit the official YOLO site →
Visual model
What a YOLO output actually looks like.
This is a stylized version of one frame after the model has run. Each box has a label and a confidence score.
Two detections, two confidences. The planner decides where to set the threshold for action.
In our pipeline
How we use YOLO for traffic-control detection.
- Sample. For each detected intersection, we pull several Street View headings facing into the crossing. This gives every approach a fair chance.
- Detect. The fine-tuned YOLO model runs on each frame and returns boxes labeled
stop_signortraffic_lightwith a confidence score. - Geolocate. Detections are projected from image coordinates back to the geographic location of the camera and the approach direction.
- Aggregate. Multiple detections of the same physical sign across nearby frames are merged into a single point feature.
- Export. The result is a GeoJSON layer of traffic-control points, each tagged with confidence and the Street View frame it came from.
What it found
Results across the pilot.
Traffic lights
0.58F1 score
9,950Street View images processed
Lights are harder. They sit far from the camera, get backlit, and disappear into tree canopy. Recall takes the hit.
Why these numbers are useful
An F1 of 0.77 isn't perfect, but for screening it's plenty. A planner reviewing a corridor can spot-check a handful of detections in the explorer and immediately see whether the model is consistent in their study area before relying on it.
Honest limits
Where YOLO stumbles.
- Backlit lights. When the sun is behind a traffic signal, the housing becomes a silhouette and the colored bulbs wash out. Recall drops sharply.
- Tree canopy. Foliage obstructing the signal head is the single biggest source of false negatives in our pilot.
- Distance and resolution. Lights mounted on far-side mast arms appear small. YOLO's bounding box gets noisy at low pixel counts.
- Look-alikes. Other red octagons (private property signs, shop displays) will be flagged occasionally. Confidence-thresholding cleans most of it up.
- Camera angle. Street View headings aren't always perfectly perpendicular to the cross-street. We sample multiple headings to compensate, but it's not free.
Practical takeaway: trust stop-sign detections more than traffic-light detections, threshold by confidence when you need precision, and always use the explorer's Street View link to verify before citing a single point in a memo.
Keep going
Now meet the other half of the pipeline.
Up next · Segmentation
U-Net 101
Where YOLO draws boxes, U-Net classifies every pixel. That's how we get crosswalk polygons from aerial imagery.
Read U-Net 101 →Try it
See YOLO outputs in the explorer
Pick a neighborhood, toggle the traffic-light and stop-sign layers, and click any detection to verify it in Street View.
Open the app →