From selected frames to a carrier event.
Tollscopic DOT# uses focused extraction and catalog scoring to turn a vehicle track into a carrier identity, a ranked candidate set, or an explicit refusal. Stronger reasoning is spent only when the evidence justifies it.
Six stages. Each one earns its cost.
Lightweight, deterministic work happens first. Heavier semantic work runs on prepared evidence, not raw video. Scoring happens against a finite catalog. Refusal is a first-class output.
Crops, contrast, denoise, normalize. The panel is prepared before any model reads it.
Each frame returns field candidates with confidence — not a final answer.
Reads are compared across frames. Agreement, disagreement, and frame diversity are surfaced.
Plausible FMCSA candidates are retrieved from every observed field.
The scorer evaluates joint evidence against candidates plus the out-of-catalog null.
Confidence class, usage policy, evidence references, version tags. The event leaves the system.
Prepare the evidence before the model reads.
Cropping, contrast, denoising, and frame normalization make the door panel more legible and reduce how much irrelevant image area the semantic extractor sees. Plain-English labels — no algorithm names on the public page.
A frame can say 'I see these possible digits' — or 'the panel is not visible.'
That uncertainty is kept. The system does not turn the first plausible string into a carrier.
{
"frame_id": "03",
"usdot": {
"digits": ["3","8","4","7","1","2","8|9"],
"confidence": [0.95,0.94,0.92,0.91,0.92,0.88,0.55]
},
"name": {
"consensus": "NORTHRIDGE FREIGHT CO",
"edit_dist": 1
},
"authority": "MC-71458",
"city_state": ["TOLEDO", "OH"],
"placard_visible": true,
"frame_quality": 0.92
} Reads are compared. Diversity matters. Duplicates don't.
Per-frame extractions are aggregated into one observation for the vehicle. Frames that look almost identical do not produce fake confidence. Frames that look different and still agree produce the most.
Which carrier explains the joint evidence best?
The catalog stage retrieves plausible carriers from all observed fields. The scorer asks whether the observed evidence is more likely if a candidate carrier is the true carrier — or if the out-of-catalog option explains it better than any carrier.
See each stage on a single track.
A schematic walkthrough of how the cascade handles three different tracks, with the verdict each pass produces at the end.
The cascade is not a demo. It is the system.
If you want to see your traffic mix go through it, we can stand up an audit-only run on selected lanes and walk through the events together.