(Deep) Learning to Fly
summary
Project builds a mono‑camera autonomous racer on a Parrot AR.Drone using FFmpeg, OpenCV, C++ and Python, with a ResNet-50 fully convolutional depth network fine‑tuned for upscaling to per‑pixel depth. Training uses structure‑from‑motion to create a pillar learning 100k frame dataset instead of manual ranges, producing metric depth after histogram normalization. Model is 200 MB but runs at ~100 ms inference on two GTX Titan X GPUs. System implements tracking, fly‑to and obstacle‑avoidance levels and succeeds in real office flights despite occasional comms glitches.
AI-written summary · about
chapters (14)
- 0:00Start
- 1:27Calibration Chess Board
- 6:35Basics
- 8:42Tracking
- 11:27Graphical User Interface
- 13:30Level Three Autonomy
- 18:47Optical Flow
- 21:08Machine Learning
- 23:15Ultimate Test
- 27:45Pillar Learning
- 32:52Max Pooling
- 37:49Fine-Tuning
- 40:43Python Application
- 41:27Ultimate Demonstration
related talks
48m(Deep) Learning to Fly
01h 00m(Deep) Learning to Fly
21mVisualizing robot minds, and other fun stuff
52mFlying services with a drone
49mUsing Deep Learning for Perception in Autonomous Systems: A Perspective from Control Theory
36mDeep Learning - Segmentation, Identification and Classification
01h 30mDeep Learning Techniques for Monocular Depth Estimation
31mDeep learning for self-driving vehicle technology