8. Deep Learning for Autonomous Driving
Training in Deep Learning
SGD Adam Weight decay
NAG AdaGrad/AdaDelta
Batch Normalization Momentum
Drop out Learning rate
Data Augmentation
Deep Learning for Computer Vision
Enhancement Denoising Object Tracking
Contour Extraction Depth Estimation
Object Detection Object Recognition
Semantic Segmentation Motion Estimation
SLAM Instance Segmentation
深度学习(Deep Learning)最近10年在AI领域的飞速发展有目共睹,
尤其在计算机视觉方面突破了以前设计图像/视频特征的瓶颈,其深度网络实现了很多应用的最佳性能
9. Deep Learning for Autonomous Driving
CNN Models in Deep Learning
AlexNet GoogleNet ResNet SSD
Faster R-CNN YOLO (1, 2, 3) Mask R-CNN
DenseNet RetinaNet MobileNet
Platforms in Deep Learning
Caffe(2) TensorFlow Torch Paddle
Keras Theano CNTK MXNet
RNN in Deep Learning
LSTM GRU
Deep Reinforcement Learning
NAS
Encoder-Decoder Model in Deep Learning
FCN SegNet ParseNet FlowNet
U-Net SFM-Net Capsule Network
GAN(Generative Adversarial Network)
深度学习的发展,除了自身机器学习理论的进步(比如GAN)以外,是跟计算机硬件的发展(Nvidia GPU)分不开的,
特别是深度学习界的开放态度和开源代码起了推波助澜的作用(如Caffe,TensorFlow)。
10. Deep Learning for Autonomous Driving
Deep Learning in Autonomous Driving
Obstacle Detection Obstacle Tracking Lane Detection
Free space Extraction Scene Segmentation Traffic Sign Detection & Recognition
Traffic Light Detection & Classification Landmark Detection
Sensor Calibration Sensor Fusion (early/late)
Vehicle Feedback Control Driving Behavior Modeling
在自动驾驶领域,我们可以看到,除了以上提到的视觉问题以外,深度学习也在多传感器标定(calibration)和融合(fusion),
以及驾驶行为和反馈控制等方面体现了强大的攻关能力。
11. Camera-based vehicle detection
F Chabot et al., “Deep MANTA: A Coarse-to-fine Many-Task Network for joint 2D and 3D vehicle analysis from monocular image”.
Deep Learning for Autonomous Driving
12. Faster R-CNN
FCL + Softmax
FCN
Deep Learning for Autonomous Driving
MultiNet: Joint classification, detection and segmentation
M Teichmann et al., “MultiNet: Real-time Joint Semantic Reasoning for Autonomous Driving”.