Aims

The aims of the course are to:

• Provide a unified view of information engineering showing how signal processing, computer vision, machine learning and control relate to one another.
• Use example applications drawn from autonomous driving to provide concrete examples of important concepts and subareas of information engineering including computer vision, machine learning and reinforcement learning
• Introduce computer vision including algorithms for 3D reconstruction, registration and object recognition
• Introduce basic concepts in inference, learning and optimisation including maximum-likelihood estimation, Bayes’ rule and gradient descent.
• Introduce basic algorithms for planning and the general area of sequential decision making / reinforcement learning

Objectives

As specific objectives, by the end of the course students should be able to:

• Provide example applications of machine perception, machine learning, and autonomous decision making systems.
• Understand the mathematical basis for perspective projection and feature detection; neural networks and parameter estimation; basic planning and reinforcement learning.
• Implement methods to solve simple computer vision and machine learning problems including object detection and segmentation and sequential decision making.

Content

A: Introduction to Autonomous Driving (1L) (A. Kendall - Guest lecturer from industry)

• The anatomy of a self-driving car (Autonomous Vehicle) with description of autonomous driving hardware (the car, sensors, interfaces and actuators) 
• Motivate the need for machine perception (computer vision), learning and decision making systems
• Important sub-problems in the data processing pipeline: object detection, localisation and mapping, prediction, planning and action
• Examples of self-driving cars

B: Machine Perception: Introduction to Computer Vision (5L) (Lecturer R. Cipolla)

• An introduction to computer vision: reconstruction, registration and recognition
• Perspective projection
• Convolution with gaussians and derivatives of gaussians to provide bandpass filters.
• Edge detection using directional filters.
• Scale-space and image pyramids for feature detection
• The SIFT feature descriptor for matching image features.
• Demonstration of state-of-the-art object detection, semantic segmentation and localisation systems
• Examples paper and class

C: Machine Learning: Introduction to Deep Learning (5L) (D. Krueger)

• Training a simple classifier: logistic regression and gradient descent
• Neural networks: Multi-layer perceptrons and back propagation
• Neural networks: convolutional neural networks
• Anatomy and training of a convolutional neural network in Tensorflow or PyTorch - network architecture, loss function, weight initialization, batch size, learning rate, epochs.
• Examples paper and class 

D: Autonomous Decision Making: Introduction to Planning and Reinforcement Learning (5L) (G. Vinnicombe and one Guest Lecture)

• Introduction to planning: Shortest path problems, value functions and dynamic programming
• Introduction to reinforcement learning: Q-learning, actor-critic methods, approximations using neural networks
• Identification of open problems
• Examples paper and class