(Deep) Learning to Fly
summary
Builds a closed system that maps a single RGB frame to a depth map and low-level controls, using OpenCV, Python, FFmpeg and a Parrot AR.Drone controlled by the single move command. Training uses ~100,000 SfM-derived depth examples (normalized 0–4 m) instead of manual labelling and fine-tunes a ResNet-15 style fully convolutional network. Runtime runs at 100 ms inference on two GTX Titan X feeding a CUDA depth server, and office tests (best of ~10–20 runs) successfully track a target and avoid windows and lamps.
AI-written summary · about
chapters (14)
- 0:00Start
- 2:36Calibration Chess Board
- 7:57Basics
- 11:00Opencv
- 11:54Optical Flow
- 14:04Collision Avoidance
- 14:40Viewing System
- 18:40Visual Odometry
- 21:08Machine Learning Approaches
- 34:16Convolution
- 35:02Max Pooling
- 39:29Structure of Convolutional Neural Network
- 42:09Initial System Design
- 43:50Ultimate Demonstration
related talks
45m(Deep) Learning to Fly
01h 00m(Deep) Learning to Fly
36mDeep Learning - Segmentation, Identification and Classification
52mFlying services with a drone
49mUsing Deep Learning for Perception in Autonomous Systems: A Perspective from Control Theory
21mVisualizing robot minds, and other fun stuff
31mDeep learning for self-driving vehicle technology
57mDrones & AI - What's all the buzz about ?