Municipal & public works / Road flooding
VisionAery Liquid Leak Detection runs on the traffic and intersection cameras a city already owns. It reports standing water appearing and growing at a sag point or an underpass to the TMC and to public works in seconds.
Blue marks a puddle. Compare the original frame with its detection overlay: the same scene, not a before-and-after flood sequence. These examples do not expand the scope of the measured benchmark; scenarios outside those tested are not covered by the published figures.
01 / From camera to crew
Rain collects at the sag point of a road, in an underpass and at a low-water crossing when the inlets and catch basins cannot keep up, and the first report is often a stranded vehicle. Most of these sites already have a camera. What they lack is someone watching it.
A sag point floods from the inlet out. The catch basin backs up, water spreads across the low lane, and the TMC hears about it from a stranded driver. The ITS camera that already covers the approach can watch it. V4 puts an interest zone on the pavement at the sag and reports standing water appearing and growing in the camera's view, staging the alarm up as it spreads. The alarm lands in the TMC's VMS and as a webhook to the dispatch or work-order system, with a snapshot, so a crew can set a lane closure before a vehicle is stranded.
The National Weather Service reports that over half of all flood-related drownings occur when a vehicle is driven into hazardous flood water, and that the location is often a road that dips down to the level of a stream channel. Seeing water pool at the sag point while it is still pooling is the window in which a lane closure prevents that.
Texas law requires drivers to comply with barriers blocking low-water crossings or flooded roadways, and Louisiana's DOTD warns that flooded roads may not be barricaded at all. FHWA lists lane submersion from flooding among the causes of roadway capacity loss and counts about 12 percent of vehicle crashes as weather-related.
Every alarm carries a timestamp and the original and overlay snapshots. When the water or the leak was detected, what it looked like and where it was are on record the moment it happened.
02 / Existing cameras. Existing systems.
Where the city's ITS or intersection cameras are Axis ARTPEC-9 (available now) or Hanwha Wisenet 9 (ships November 15, 2026), the app installs on the camera. No server at the intersection, and nothing leaves the cabinet but the alarm. A PTZ on a tour carries its own interest zones and exclusion zones at each preset.
Where the TMC has a mixed camera fleet, a server takes RTSP streams from the cameras already on the network: an Emerson PACEdge industrial PC, an x86 server with an integrated Intel GPU or an NVIDIA GPU, or an NVIDIA Jetson, all shipping November 15, 2026. A fanless Jetson suits a roadside cabinet; a rack server in the TMC suits the camera fleet; the stream count per box is sized at quote.
On-camera, native camera events go into the VMS the TMC already runs. On-server, alarm triggers go into Nx Witness, Wisenet WAVE, Milestone and AXIS Camera Station, with an overlay RTSP stream per channel. MQTT as JSON or Sparkplug B, pre-mapped for Ignition, into the utility's SCADA, and into DeltaV over MQTT on Emerson. HTTPS webhooks to email, SMS, ticketing or a BMS.
On a road low point the point is to watch during the rain, so the gate is left off there and the alarm stages up only when a detected pool grows and persists.
03 / Know the limits
The road-flooding image is a visual example, not an extension of the measured benchmark scope.
Measured on 9,147 previously unseen frames, including true positives and negatives, out of the box with no per-site training, before the alarm engine is applied.
| F-score | Detection rate | False positive rate | |
|---|---|---|---|
| Puddle | 0.889 | 97.1% | 0.8% |
| Spray and visible vapor | 0.923 | 98.1% | 3.7% |
F-score is the pixel-level harmonic mean of precision and recall. Frame-level figures; the alarm engine's staging, persistence and color-confidence checks exist so that a false frame does not become a false alarm.
Scenarios outside those tested are not covered by these figures.
Detection within seconds · Installed and configured in under 30 minutes · About 25 pixels per foot on target, nothing else
Detection depends on the view, lighting and image detail. VisionAery Liquid Leak Detection does not identify chemical composition, detect invisible gas or measure concentration, depth or volume. It supports, not replaces, inspections, required instrumentation, cleanup protocols and safety procedures.
Questions
