Problem
Support rail-yard operations by detecting, identifying, and tracking rail assets across real-world camera views.
Context
The source portfolio describes a multi-camera system with edge and control-center components deployed around NVIDIA AGX Xavier hardware.
My role
Built computer vision analytics, processed datasets, and supported edge deployment.
Constraints
- Outdoor industrial conditions and long camera ranges.
- Multi-camera identity continuity.
- Edge compute and operational reliability.
Architecture
Detection and tracking ran within an edge-to-control-center design, with re-identification supporting continuity across views.
Technical decisions
- Used specialized detection, tracking, and re-identification stages.
- Deployed on NVIDIA edge hardware for field operation.
Trade-offs
The system design balanced model accuracy, latency, bandwidth, edge compute limits, outdoor camera conditions, and operational reliability.
Results
This work contributed to an industrial computer-vision research effort and a peer-reviewed publication on railcar detection, identification, and tracking.
Public note
Field screenshots and deployment diagrams are omitted from the public portfolio because operational context can be sensitive.
Related links
- University at Albany project announcement
- Chang, M.-C., Zhao, G., Pandey, A. K., Pulver, A., and Tu, P. “Railcar Detection, Identification and Tracking for Rail Yard Management.” IEEE ICIP, 2020.