Project Overview
Led the development and deployment of an AI-powered video analytics system for public transport that
counts passengers and detects vehicle occupancy in real-time. The system processes video feeds from
onboard cameras using computer vision and deep learning models to provide accurate, GDPR-compliant
passenger data for transit operators across Europe.
The Challenge
Public transport operators needed accurate passenger counting data to optimize routes, improve
service quality, and comply with COVID-19 capacity regulations. Traditional manual counting methods
were expensive, inaccurate, and impossible to scale. The solution had to:
- Work in challenging real-world conditions (varying lighting, crowding, occlusions)
- Meet strict GDPR requirements for passenger privacy
- Achieve >90% accuracy to be operationally viable
- Run on edge devices for real-time processing
- Integrate with existing transport management systems
My Role
As AI Product Manager and Technical Lead, I:
- Led the 4-person AI/ML team through the complete product development lifecycle
from research to production deployment
- Defined product strategy and roadmap based on customer interviews and market
research across 5+ European transport operators
- Secured €100K in public funding from German research grants (ZIM program) for
AI innovation
- Established GDPR-compliant data frameworks working with Legal, Engineering,
and data protection officers to enable legal customer data collection
- Coordinated cross-functional teams across Engineering, Sales, QA, and
regulatory compliance
- Managed model training and optimization including dataset creation, annotation
workflows, and accuracy validation
Solution & Technical Approach
We developed an end-to-end computer vision pipeline:
Technology Stack
Python
PyTorch
OpenCV
YOLO
TensorRT
Docker
Edge Computing
MQTT
Key Technical Components
- Object Detection: Custom-trained YOLO models for passenger detection optimized
for edge deployment
- Tracking Algorithm: Multi-object tracking to count passengers entering/exiting
through vehicle doors
- Privacy-by-Design: On-device processing with no raw video storage, only
anonymized counting data transmitted
- Edge Optimization: Model quantization and TensorRT optimization for real-time
inference on NVIDIA Jetson devices
- Data Pipeline: Automated annotation workflows and continuous model retraining
pipeline
