uSAR: UAV Edge Human Detection for Search & Rescue

Overview

uSAR is a UAV-based human detection system for search-and-rescue in disaster management. Drones are well-suited to disaster relief — small, mobile, and able to cover terrain quickly — so automatic person detection in drone imagery can meaningfully speed up response. The system detects people from aerial views and is optimized to run on-edge for real-time inference rather than offloading to a server.

Approach

• Built on the SARD dataset — people across diverse terrain and backgrounds, labelled by activity (running, walking, standing, sitting, lying), split 1,534 train / 387 test.
• Fine-tuned a pre-trained EfficientDet-Lite-4 detector — an EfficientNet-Lite backbone with a BiFPN feature extractor and shared box predictor.
• Exported to TensorFlow Lite, optimized for real-time inference on mobile CPU/GPU and EdgeTPU.