Trends in Machine Learning for Unmanned Aerial Vehicle Applications - 鄭聖文

歡迎來到 MOPCON 2024 共筆

共筆入口：https://hackmd.io/@mopcon/2024
手機版請點選上方 按鈕展開議程列表。

從這開始

• 投影片連結

• Introduction to Artificial Neural Network

  • Neuron (perceptron)
    • book: hands-on convolutional neural with tensorflow(
  • (Deep) Neural network (multi-layer perceptron)

    是一種通用函式擬合器，可以用來模擬任何物理現象的模型。

    • book: data-driven science and engineering
    • 可以較輕鬆將類神經網路用線性代數表達。
    • Softmax function
      
    • Training a Neural Network (Back Propagation)
      • How to Train an Artifical Neural Network(Prof. Youn-Long Lin)
    • Residual Network (ResNet) 2015-
    • Dropout
    • Encoder-Decoder/AutoEncoder

      萃取後，再重建。

• Object Detection and Segmentation

  • Object tracking applications of drones
    • Ukraine War
    • Agriculture
    • Firefighting drones in China
    • Search and Rescue
    • Border Patrol
    • Bridge Inspection
  • Feature extraction of convolutional neural network (CNN)
    • low level feature: edges, dark, spots
    • mid level feature: eyes, ears, nose
    • high level feature: facial structure
  • Max pooling
    • Pooling in Convolutional Neural Networks
  • Example: VGG16 network for image classification
    • VGG16 (Visual Geometry Group Architecture 16), University of Oxford 2014
      
  • Object Detection (YOLO v1)
    • YOLO, You Only Look Once
  • Fast R-CNN VGG-16 vs CNN
    • mAP(mean Average Precision): 73.2% vs 63.4%
    • FPS: 7 vs 45
  • Semantic segmentation: segNet
    • Auto-Driven vehicle
  • Semantic segmentation: U-Net

• Learn to control: Reinforcement Learning

  • Reinforcement Learning (RL)
    • There is an Agent that interacts with the environment
    • value function
      $V$ and quality function
      $Q$
  • Deep Reinforcement Learning
    • 傳統而言，以力學與數學進行分析與控制。
    • With Deep Reinforcement Learning
      • Control of a Quadrotor with Reinforcement Learning, 2017
    • Champion-level flight of Quadrotor with Reinforcement Learning
    • Collision-free Flight with Reinforcement Learning
    • Digial Twin
      • Software simulators for robots

• Transformer Model

  • transformer and attention model
  • Encoder-Decoder architecture
  • Embeddings are vectors in high dimensions
  • variants
    • Encoder-only
    • Decoder-only
    • Encoder-Decoder
  • $Attention(Q,K,V)$
  • Self-attention vs Cross-attention

• Vision Transformer

  • Model generalization
  • Visual Transformer (ViT)