The rest of this piece is what that means, why it works this way, and what it implies for the buyer who needs the data.
Why it matters
Detection in the cloud has one architecture: the vehicle uploads full-resolution video, and a server runs a model on it. Bandwidth, latency and cost all scale with the raw video volume. A network of cameras streaming video for cloud inference is a network paying twice, in cellular data and in server compute.
Edge inference inverts the trade. The camera runs the model, uploads metadata, and only sends the image up when it clears a quality threshold. What the customer sees is a detection stream, not a video stream.
What edge inference is
Edge inference means running an AI model on the device where the data is captured. In this case, the camera. Empiric Earth CityStream turns what road-facing cameras see into data products for public-sector and mapping customers, and the "sees" verb depends on the model running on the camera.
Both consumer and fleet cameras in the sensor network run the same edge inference. That is a network-wide capability, not a premium tier.
How the measurement works
A detection appears within a couple of seconds of the camera seeing it. The image follows in a couple of minutes once the metadata clears the quality thresholds required for upload.
Every detection is indexed into H3 hexagons, the standard grid that divides the map into equal hexagonal cells, and map-matched to an OpenStreetMap node so it lands on the nearest road. Repeat sightings of the same object aggregate at H3 resolution 12.
Detection thresholds are set per object class, using the precision-recall trade-off measured when the model was trained. Customers can move the confidence filter themselves: raise the confidence and see fewer detections, lower it and see more.
Access is through a web interface and through APIs, so the same live feed drives a dashboard for a traffic engineer or a downstream pipeline for a mapping team.
Where Empiric Earth fits
The live view doubles as the clearest demonstration of what the network covers. What a customer sees is a map of real-time detections across the network, described honestly as a sample of the data rather than the whole of it.
Detection density is highest in New York City in current demos. That is where the network runs densest today, and that fact is a coverage statement more than a marketing one: what the live map shows in Manhattan is not what the live map shows in a rural county.
Every detection ships with the image that produced it, and mapping teams use that imagery to update their own map data after a construction project or another change alters a road. The imagery is the audit trail. It is not an inference about the world; it is what the camera saw.
Only we know
The default retention on each object class is a measurement decision, not a storage policy. A work zone object has a lifetime of 24 hours and vanishes if no camera sees it again. A road sign holds for 30 days or longer. A detection stays live only while cameras keep re-seeing it, and refreshes each time they do.
Setting those clocks required real observation of how long each class of object usually persists on a road. That distribution is not publishable data. It is only knowable to a company that runs the same measurement across its sensor network on a schedule that catches both the arrival and the disappearance.
Limits
Curated products run on a weekly cadence, alongside the live view. The curated traffic-sign product returns better sign position and better sign meaning than a raw live detection does, because it aggregates many sightings before it decides. The two speeds are complementary, not interchangeable: a live view answers "what is on the road right now"; a weekly curated view answers "what is the road usually like."
The Work Zones product assembles scattered object detections into the road segments a work zone actually affects, instead of leaving isolated pins for the customer to interpret. That aggregation lives on the curated side, because it needs multiple sightings.
FAQ
Q: What is edge inference? A: Running the AI model on the camera itself, rather than uploading raw video to a cloud service that runs the model.
Q: How fast does a CityStream detection appear? A: A couple of seconds from being seen, with the image following in a couple of minutes.
Q: Do consumer and fleet cameras both run edge inference? A: Yes. Both run the model at the edge.
Q: How is a detection placed on the map? A: Indexed into H3 hexagons and map-matched to an OpenStreetMap node. Repeats aggregate at H3 resolution 12.
Q: How does CityStream decide when a work zone is gone? A: A work zone object defaults to 24 hours from the last sighting. When the network stops seeing it, it disappears from the live view.