Ensuring robust vessel detection under varying weather and lighting conditions
Refactored legacy codebase and modularized key components for maintainability
A harbor-monitoring computer vision system for detecting vessel type and direction from real Copenhagen camera streams.
Context
The project sat at the intersection of urban operations and messy real-world video. Harbor cameras see weather, occlusion, reflections, and many vessel types, so the task was less about a pristine model demo and more about making detection useful under practical deployment constraints.
Build
I worked on the model pipeline, training flow, and deployment structure around PyTorch, AzureML, Kubernetes, and CI/CD. The system had to classify vessel types and sailing direction while remaining maintainable enough for future retraining and broader rollout.
Engineering
Refactored legacy codebase and modularized key components for maintainability
Trained custom computer vision models using PyTorch to detect vessel types and motion
Integrated with AzureML for cloud-based training and experimentation
Streamlined Kubernetes deployment with improved CI/CD setup for reliable retraining and rollout
Outcome
Stack