01. Problem Context & Conflation Bottleneck
In large-scale digital mapping, update velocity at scale is critical. HERE UniMap processes 500 million kilometers of probe data hourly to generate a dynamic digital representation of roadway reality. However, the ingestion pipeline faced an operational bottleneck in Spatial Conflation: programmatically associating detected traffic signs with their specific physical mounting structures (e.g., overhead gantries, cantilever poles, ground posts).
The Legacy Performance Ceiling
The baseline XGBoost model (F1 score: 0.85) produced an approximate 14% association error rate. When spatial proximity was ambiguous, the system required human verification or physical 'HERE True Drives' to inspect physical roadway infrastructure.
98% Accuracy
Peak model performance
1,500+ Hours
Weekly field capacity saved
28 Features
Custom engineered for Lidar
02. Neural Architecture
I replaced the legacy tree-based model with a two-tower neural network designed for high-order spatial reasoning. The architecture independently encodes the geometric attributes of mounting structures ('Parents') and detected sign assets ('Children').
Parent & Child Tracking
Encodes spatial properties of massive mounting structures (12-foot highway gantries) and small traffic assets (30-cm urban signs).
Joint Spatial Reasoning
Processes 28 engineered features, prioritizing Bearing Alignment and Altitude Ratios to override noisy Lidar coordinates.
Structural Link Prediction
Achieved 98% accuracy in identifying the correct physical mounting structures for traffic assets.
03. Failure Mode Analysis: Hitting the 'Sensor Floor'
At 98% accuracy, detailed failure analysis revealed that remaining errors were no longer model misclassifications, but upstream sensor artifacts and ground-truth labeling errors:
1. LiDAR Sensor Artifacts
The network correctly identified instances where LiDAR sensors vertically compressed stacked signs or generated positioning jitter across highway junctions.
2. Ground Truth Omissions
The model identified valid mounting links that human annotators had historically omitted in ground-truth labels.
Evolution to Automated Audit
These diagnostics enabled the model to function not just as an association classifier, but as an automated data validation step flagging upstream sensor defects within the UniMap pipeline.
04. Field Impact & Pipeline Performance
Fieldwork Reduction Potential
By automating mounting structure association with 98% accuracy, the model established the foundation to eliminate an estimated 1,500 hours per week of manual verification drives across European road networks.
Downstream Navigation Integrity
Minimizing structural misclassifications lowered spatial error propagation for commercial OEM navigation partners including BMW and Mercedes-Benz.