On-device disaster scene parsing

Our team built Hazard View Bird, an on-device disaster scene parsing system to help first responders assess damage from aerial footage. We used transfer learning with EfficientNet-B0 to segment and classify 14 types of disaster damage in real time.

I worked on optimization for edge deployment: pruning, quantization, and ONNX export for low-power hardware like NVIDIA Jetson, cutting processing time by 40%. The project placed 3rd at the Khoury College Annual Project Pitch-A-thon.

Focus

• On-device computer vision for disaster response
• EfficientNet segmentation with edge optimization
• ONNX deployment on Jetson-class hardware