UAV-based Person Detection and Geolocation for Search and Rescue
summary
Pipeline builds a custom UAV dataset of about 2,000 images captured with a Phantom 5, heavily augmented to simulate snow rain and occlusion, then retrains YOLO variants with transfer learning to raise average precision from ~40% (off‑the‑shelf) to 95% AP on the test set. An offline geolocation step casts rays against a digital elevation map to place detections on the ground with roughly 10 m error. Infrared experiments fail in extreme seasons, and the system is already used by rescue teams.
AI-written summary · about
chapters (11)
- 0:00UAV person detection intro
- 1:13Search and rescue motivation
- 3:46Drones in SAR missions
- 7:20Building drone dataset
- 10:41Real-time detection model
- 13:11Data augmentation & performance
- 15:50Geolocation of persons
- 18:43Position from drone data
- 20:50Tracking movement & direction
- 22:17Future work & infrared
- 23:00Q&A
related talks
52mFlying services with a drone
21mUsing Computer Vision to Solve Interesting Problems
36mDeep Learning - Segmentation, Identification and Classification
01h 14mFoundations of Object Detection from an Expert's Overview
10mLightning Talk: Ball Tracking and Detection in Soccer Videos - Comparison of...
24mAI in maritime - Autonomous ships
13mFlyMeThrough: Human-AI Collaborative 3D Indoor Mapping with Commodity Drones
47mNext-gen Dynamic UAV using Power9 Systems