Frames analysed (logged)
0
Frames flagged waterlogged
0
Incidents opened
0
The benchmark runs the exact detector used by the live monitor over 26 held-out, openly licensed photographs (13 waterlogged, 13 hard negatives: dry roads, rain-wet asphalt, night reflections). Every number below is measured in this browser at run time — nothing is hard-coded.
Recorded runs in the model registry
What this system does and does not claim
- Stage 1, the in-browser detector, is a deterministic HSV + texture-gradient estimator with a temporal persistence gate — not a trained segmentation network. It is tuned for recall and used as a cheap screening filter: on hard negatives (rain-wet asphalt, night reflections, bright dry concrete) it over-flags, which the measured stage-1 false-alert rate above reports openly.
- Stage 2 is the AI frame verification: only screened-in frames are sent to the vision model, and its verdict decides the alert. The end-to-end precision, recall, F1 and false-alert rate shown above are the numbers an operator actually experiences. People, vehicle and depth figures come from this stage only.
- Severity is a documented weighted formula; every contribution is shown in the live monitor so an operator can audit a score.
- Night, heavy rain, glare and freshly wet asphalt remain the dominant error sources. Raise the detection threshold for glare-prone cameras.