HERE Berlin Hackathon 2026 · April 10

School Zone Speed Limit
Validation

Confidence Layer for the Hard 20%

Potsdam & South-West Berlin | Partnered with Siemens Digital Logistics

HERE already handles 80%

44 million connected vehicles detect road signs weekly.
The automation pipeline handles the clear-cut cases well.

We built the confidence layer for the hard 20%.

Three Deliverables

✅

Valid Zone?

7-gate deterministic classification with visual evidence

📍

Zone Extent

Graph-walk to terminating sign along road network

🕐

Temporal

Permanent vs conditional with OCR time extraction

Six Datasets — 48 Million+ Records

Dataset	Size	Purpose
School zone signs	326 rows	Classification input
Remaining speed signs	11,607 rows	Extent end-sign detection
Road geometry	27,943 links	Graph-walk topology
School POIs (4,780 → 3,985 educational)	3,985	Proximity gate G2
Mapillary images	1,850 imgs	Visual evidence + OCR
Probe speed data	48.5M records	Speed anomaly signal

Deliverable 1: Is It a Valid School Zone?

7 deterministic gates — no LLM in the decision loop

G1 Tempo-30-Zone Exclusion

G2 School Proximity (OSM + POI + Geoportal)

G3 Florence-2 Visual Detection Score

G4 Temporal Classification (OCR regex)

G5 Detection Confidence (n_detections)

G6 Spatial Consistency (cross-pole)

G7 Crop OCR Final Validation

Gate Classification Results

219

AUTO_APPROVE

NEEDS REVIEW

EXCLUDE

80 excluded because they sit inside an existing Tempo-30-Zone (G1)
— the zone already governs the speed limit, sign is advisory only

Beyond Poles — Zone-Level Confidence

DBSCAN clustering (ε=300m, UTM EPSG:25833) groups poles into school zones

Spatial clusters

Matched to a school
within 500m

CONFIRMED zones

CONFIRMED 33 MIXED 16 NEEDS_REVIEW 12 EXCLUDED 9

Deliverable 2: Where Does the Zone End?

Map-match pole to road link, walk forward in bearing-governed direction, stop at terminating sign.

208

Extent LineStrings
Mean length: 224 m

Termination types:

END_SIGN

77 (37%)

DEAD_END

57 (27%)

ROAD_END

49 (24%)

MIRROR

17 (8%)

MAX_DIST

8 (4%)

Deliverable 3: Permanent or Conditional?

57.1%

Permanent
Active always, incl. holidays

42.9%

Conditional
Time/day restrictions

Method: Florence-2 detect → crop → Tesseract OCR → regex

Step 1: Florence-2 Detects the Sign

Raw Mapillary street-view image → bounding boxes drawn around every sign

Florence-2 detection with green bounding boxes on pole 00250

Pole 00250 — 30 km/h sign + supplemental time plate detected with bounding boxes

Step 2: Crop → OCR → Time Restriction

Bounding box crop — sign fills the frame

        OCR reads:

        "30"

        "Mo-Fr 6-18 h"

CONDITIONAL

Active Monday–Friday, 06:00–18:00

School Warning Sign — Pole 235

Florence-2 detects both Zeichen 136 (school crossing) and the 30 km/h plate

Full scene with bounding boxes

Cropped — clear school crossing sign

Why Not Just Ask a VLM?

Model	Method	Hallucination Rate
Gemini 2.5 Flash	Direct image query	89.8%
GPT-4o	Direct image query	72.3%
Claude 3.5 Sonnet	Direct image query	61.4%
Florence-2 + CoT prompt	Detect → crop → LLM	17.9%
Florence-2 + Our Pipeline	Detect → crop → Tesseract	5.4% ✓

Signs are <5% of frame. Crop first, then read. CoT prompting cut hallucination by 80%.

Live Demo

315 poles · extent LineStrings · DBSCAN zone clusters — all interactive

What We Delivered

315

Unique poles analysed

219

Auto-approved zones

Confirmed zone clusters

208

Extent LineStrings
mean 224m

5.4%

False positive rate
(vs 89.8% for direct VLM)

VLMs benchmarked
3 ML approaches

Deterministic.

Threshold-tunable.

Production-ready.

No LLM in the decision loop · Multi-source triangulation · QA-ranked output

School Zone Speed LimitValidation