DATA SCIENCE INTERN @ HERE TECHNOLOGIESAUTONOMOUS MAPPING & ML

Autonomous Mapping:
Neural Spatial Conflation

Deployed a predictive neural network that improved road-sign detection accuracy from 82% to 98%, eliminating ~1,500 manual fieldwork hours weekly and lowering European data collection costs.

01. Executive Summary

In the race for global mapping leadership, update velocity at scale is the primary differentiator. HERE UniMap addresses this by processing 500 million kilometers of probe data hourly to create a unified digital twin of reality. However, the system faced a critical operational bottleneck in Spatial Conflation: the ability to programmatically link traffic assets (e.g., speed limit signs) to their specific physical mounting structures.

The Operational Ceiling

The legacy XGBoost model (F1 score: 0.85) created a ~14% error tax, requiring human audits to verify associations. When data was ambiguous, the business relied on expensive physical "HERE True Drives" to verify reality on the ground.

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 architecture with a Neural Network designed for high-order spatial reasoning. Unlike standard classifiers, this model independently encodes the geometry of "Parent" structures and "Child" signs.

Parent Geometry

Mounting structures (e.g., 12-foot highway gantries)

Child Geometry

Traffic assets (e.g., 30-cm urban signs)

Neural Network

Joint Spatial Reasoning

Processes 28 engineered features, prioritizing Bearing Alignment and Altitude Ratios to override noisy Lidar coordinates.

Output

Structural Link Prediction

98% Accuracy

03. Hitting the "Sensor Floor"

The most significant outcome of this initiative was the discovery that at 98% accuracy, the model had hit the "Sensor Floor." By interrogating the remaining error rate, I proved that the failures were no longer logical, but were caused by upstream detection issues:

1. Lidar Hallucinations

The model correctly identified when sensors vertically "pancaked" stacked signs or teleported objects across junctions.

2. Ground Truth Omissions

The model flagged valid structural links that human annotators had historically missed.

Evolution to "Automated Auditor"

This insight elevates the model from a simple automation tool to an Automated Auditor, allowing the team to flag sensor failures in real-time and improving the overall integrity of the UniMap pipeline.

04. Business Impact

Capacity Redeployment

By automating the association logic, we move closer to eliminating the need for manual verification drives, saving an estimated 1,500 hours per week pan-Europe.

Enterprise Risk Mitigation

Reducing misclassifications lowers the risk of downstream navigation errors for OEM partners like BMW and Mercedes.