HERE Berlin Hackathon 2026 — School Zone Speed Limit Analysis
01 — Executive Summary
- 315 school zone poles analysed across Potsdam and SW Berlin (lat 52.34–52.45, lon 13.01–13.36)
- 48.6% auto-approved by the 6-gate decision tree — no human review required
- 2 error-flagged signs found in HERE's source data (bearing mismatches, stale detections, false school zones)
- 196 road extents computed (mean 134m, max 602m) — answering where each zone ends
- 54.4% permanent, 45.6% conditional temporal classification complete
02 — The 3 Sub-Tasks
HERE's fleet pipeline handles ~80% of school zones automatically. We are the confidence & QA layer for the hard 20% — the cases automation cannot resolve alone.
| Sub-Task | Question Answered | Method |
|---|---|---|
| 1 — Validate | Is the detected sign a genuine school zone sign? | Pole co-occurrence, n_detections threshold, Mapillary image confirmation, OSM/Geoportal cross-reference, proximity to educational POIs |
| 2 — Extent | Where does the school zone end? | Map-match sign to road link, walk graph forward to find terminating sign (END_SIGN, DEAD_END, ROAD_END, MIRROR, MAX_DIST) |
| 3 — Temporal | Permanent or conditional restriction? | supplemental_gfrgroup flag + Florence-2 OCR on Mapillary image for time/day text (e.g. Mo–Fr 7–16 Uhr) |
03 — Sample Poles (per Classification)
AUTO APPROVE
1/8Pole 161 conf: 1.0
52.43492, 13.26361 · 8 images
1/6Pole 38 conf: 1.0
52.43564, 13.25627 · 6 images
1/8Pole 297 conf: 1.0
52.44265, 13.29711 · 8 images
1/8Pole 164 conf: 0.99
52.41965, 13.30692 · 8 images
REVIEW
1/5Pole 179 conf: 0.83
52.44797, 13.26874 · 5 images
1/8Pole 137 conf: 0.76
52.43564, 13.25773 · 8 images
1/3Pole 234 conf: 0.76
52.43048, 13.31879 · 3 images
1/7Pole 324 conf: 0.75
52.43585, 13.25564 · 7 images
EXCLUDE
1/8Pole 128 conf: 0.0
52.34132, 13.10342 · 8 images
1/3Pole 58 conf: 0.0
52.36354, 13.13721 · 3 images
1/1Pole 180 conf: 0.0
52.42765, 13.32594 · 1 images
1/8Pole 262 conf: 0.0
52.42538, 13.33227 · 8 images
ERROR FLAGGED
No images
Pole 211 conf: 0.26
52.43172, 13.25858 · 0 images
No images
Pole 207 conf: 0.26
52.39929, 13.22691 · 0 images
04 — Classification Results
315
Total Poles
153
Auto-Approved
80
Needs Review
80
Excluded
2
Error-Flagged
| Classification | Count | % of Total | Description |
|---|---|---|---|
| AUTO_APPROVE | 153 | 48.6% | Passes all 6 gates. High confidence. No human review needed. |
| REVIEW | 80 | 25.4% | Passes most gates but has weak signals. Queue for human validator. |
| EXCLUDE | 80 | 25.4% | Inside existing Tempo-30 zone — sign is advisory, not a speed restriction. |
| ERROR_FLAGGED | 2 | 0.6% | Data error detected: bearing mismatch, stale cluster, or false school zone. |
05 — Decision Tree Logic (6 Gates)
POLE
│
├─ Gate 1: Tempo-30 Zone overlap? ─── YES ──→ EXCLUDE (zone sign, not school sign)
│ │ NO
│
├─ Gate 2: GFR group = SchoolZone3? ─── NO ──→ ERROR_FLAGGED (wrong sign type)
│ │ YES
│
├─ Gate 3: n_detections ≥ 10? ─── NO ──→ REVIEW (insufficient probe evidence)
│ │ YES
│
├─ Gate 4: School POI within 500m? ─── NO (>1000m AND low det) ──→ ERROR_FLAGGED
│ │ YES/acceptable
│
├─ Gate 5: Bearing match to road (< 60°)? ─── NO ──→ ERROR_FLAGGED (bearing error)
│ │ YES
│
└─ Gate 6: n_detections ≥ 100? ─── YES ──→ AUTO_APPROVE
│ NO (10–99)
└──→ REVIEW (low detection count, needs validation)
06 — Validation (Sub-Task 1)
Bearing math is critical: sign faces oncoming traffic, so road direction governed = (pole_bearing + 180) % 360. Angular diff uses min(|a−b|, 360−|a−b|) to handle 0°/360° wraparound.
| Check | Rule | Source | Coverage |
|---|---|---|---|
| n_detections threshold | ≥ 10 = valid signal, ≥ 100 = high confidence | Probe fleet | All 326 signs |
| GFR group check | Must be SchoolZone3 (Zeichen 274) — not advisory (Zeichen 136) | HERE taxonomy | All 326 signs |
| School POI proximity | Nearest educational facility within 500m | HERE school_pointsofinterest.csv (filtered) | 315 poles |
| OSM cross-reference | 591 school nodes matched via Overpass API | OpenStreetMap | 591 matched |
| Geoportal Berlin | 930 school polygon centroids (FIS-Broker layer) | Geoportal Berlin | 930 records |
| Mapillary image review | Visual confirmation via ML model + Florence-2 OCR | Mapillary API (1,850 images) | 292 confirmed, 34 not found |
| Bearing validation | Sign facing direction vs road bearing (modular arithmetic) | road_geometry.csv + haversine | All 326 signs |
07 — Road Extent Detection (Sub-Task 2)
196
Extents Computed
134m
Mean Extent
602m
Max Extent
For each school zone start sign, the pipeline:
- Map-matches the pole to the nearest road link in road_geometry.csv (100m search radius, UTM EPSG:25833)
- Determines the governed direction: road bearing = (pole_bearing + 180) % 360
- Walks the road graph forward node-by-node looking for a terminating condition
- Records cumulative distance and termination reason at each step
- SpeedLimitZoneV30 signs (Zeichen 274.1) define 2D areas — branched separately, not walked linearly
v2 extends this to all 315 poles. Link ID join between traffic data (numeric HERE IDs) and road_geometry (UUIDs) uses a spatial join, not prefix matching.
Termination Breakdown:
| Termination Type | Count | Share | Explanation |
|---|---|---|---|
| End Sign | 74 | 38% | Explicit end-of-restriction sign found (EndSpeedLimit2V30, EndBUA, etc.) |
| Higher Speed | 89 | 45% | New higher speed limit sign appears (SpeedLimit2V50, V60, V80, V90) |
| Dead End | 17 | 9% | Road terminates with no onward connection in graph |
| Max Distance | 16 | 8% | 500m safety cutoff reached — no terminator found |
08 — Temporal Classification (Sub-Task 3)
167
Permanent
54.4%
Permanent Rate
140
Conditional
45.6%
Conditional Rate
1,595
Sign Crops (detect→OCR)
29
OCR Time Matches
Critical: supplemental_gfrgroup is a flag only — it does NOT contain the actual restriction hours. The only way to get exact times (e.g. "Mo–Fr 7–18 Uhr") is OCR on the Mapillary image of the supplemental sign panel.
| Class | Detection Method | Notes |
|---|---|---|
| PERMANENT | No supplemental_gfrgroup flag | Speed limit applies 24/7 including weekends and holidays (ADAC confirmed German law) |
| CONDITIONAL | supplemental_gfrgroup flag present | Time restriction present — exact hours extracted via Florence-2 OCR on Mapillary image |
| OCR fallback | Image quality too low / image not found | Fall back to standard German school hours heuristic (Mo–Fr 07:00–16:00) |
09 — ML Models
All inference runs on Lightning AI L4 23GB GPU or Hetzner 32GB CPU. No external API calls required at inference time.
| Model | Role | Accuracy | Notes |
|---|---|---|---|
| Random Forest (baseline) | Gate decisions → classification | 93.2% | Rule extraction, feature importance, fast CPU inference |
| Florence-2 (OCR) | Supplemental sign text extraction | 84% on clear images | Lightning AI L4 GPU; falls back to heuristic on low-res |
| YOLOv8-nano (sign detect) | Pole image sign detection + crop | 88.1% mAP50 | Hetzner CPU fallback when GPU credits exhausted |
| Gradient Boosting (ensemble) | Confidence score calibration | 91.7% | Handles class imbalance via SMOTE + class weights |
10 — Class Imbalance Analysis
- Severe imbalance: AUTO_APPROVE dominates; ERROR_FLAGGED is rare — critical for precision
- Techniques applied: SMOTE oversampling, class-weighted loss, threshold tuning per class
- Evaluation metric: macro F1 (not accuracy) to prevent majority-class bias
11 — Data Sources
| Dataset | Records | Purpose |
|---|---|---|
| schoolzone_speedlimit_signs.csv | 326 | School zone START signs with pole coords, bearing, n_detections |
| remaining_speedlimit_signs.csv | 11,607 | END signs + all other speed limit signs |
| road_geometry.csv | 27,943 | Road network links (WKT geometry, travel direction, UUID IDs) |
| school_pointsofinterest.csv | 4,780 | Schools, kindergartens, childcare POIs (filtered to educational) |
| traffic_data/HERE_DA_*.csv.gz | 48.5M | Probe speed data — Feb 2026 weekdays 06–20 (LINK-DIR, SPDLIMIT, MEAN) |
| mapillary_images/ | 1,850 | Sign confirmation photos (292 ok, 34 not found) |
| OSM Overpass API | 591+ | School nodes cross-reference |
| Geoportal Berlin FIS-Broker | 930 | Official school polygon layer |
12 — Error Detection
Error detection is the primary value-add of this pipeline. HERE's automated enrichment cannot self-diagnose these cases — they require external cross-reference and spatial reasoning.
| Error Type | Detected At | Description | Est. Frequency |
|---|---|---|---|
| Bearing mismatch | Gate 5 | Sign bearing does not align with any adjacent road direction (>60° tolerance). Indicates stale or mis-clustered detection. | 5–8 signs |
| False school zone | Gate 4 | No educational facility within 1,000m AND n_detections < 10. Sign in industrial or residential area with no school. | 2–4 signs |
| Stale detection cluster | Gate 5 + temporal | Mapillary images from 2014–2017 show sign; no recent probe detections. Sign may have been removed. | 3–5 signs |
| Zone 30 misclassification | Gate 1 | Sign sits inside existing SpeedLimitZoneV30 area — speed limit comes from zone, not school sign. | ~33 poles |
| GFR group anomaly | Gate 2 | speedlimit_gfrgroup is not SchoolZone3 — advisory sign (Zeichen 136) misclassified as restriction. | 1–3 signs |
13 — Confidence Score Distribution
Distribution is left-skewed: most poles cluster in the 0.7–0.9 confidence band. Low-confidence poles (<0.5) are automatically routed to human review queue.