Autonomous Cars: Deep Learning and Computer Vision in Python
Autonomous Cars: Computer Vision and Deep Learning
The automotive industry is experiencing a paradigm shift from conventional, human-driven vehicles into self-driving, artificial intelligence-powered vehicles. Self-driving vehicles offer a safe, efficient, and cost effective solution that will dramatically redefine the future of human mobility. Self-driving cars are expected to save over half a million lives and generate enormous economic opportunities in excess of $1 trillion dollars by 2035. The automotive industry is on a billion-dollar quest to deploy the most technologically advanced vehicles on the road.
As the world advances towards a driverless future, the need for experienced engineers and researchers in this emerging new field has never been more crucial.
The purpose of this course is to provide students with knowledge of key aspects of design and development of self-driving vehicles. The course provides students with practical experience in various self-driving vehicles concepts such as machine learning and computer vision. Concepts such as lane detection, traffic sign classification, vehicle/object detection, artificial intelligence, and deep learning will be presented. The course is targeted towards students wanting to gain a fundamental understanding of self-driving vehicles control. Basic knowledge of programming is recommended. However, these topics will be extensively covered during early course lectures; therefore, the course has no prerequisites, and is open to any student with basic programming knowledge. Students who enroll in this self-driving car course will master driverless car technologies that are going to reshape the future of transportation.
Tools and algorithms we’ll cover include:
Deep Learning and Artificial Neural Networks
Convolutional Neural Networks
HOG feature extraction
SIFT, SURF, FAST, and ORB
Tensorflow and Keras
Linear regression and logistic regression
Support Vector Machines
Your instructors are Dr. Ryan Ahmed with a PhD in engineering focusing on electric vehicle control systems, and Frank Kane, who spent 9 years at Amazon specializing in machine learning. Together, Frank and Dr. Ahmed have taught over 200,000 students around the world on Udemy alone.
Students of our popular course, “Data Science, Deep Learning, and Machine Learning with Python” may find some of the topics to be a review of what was covered there, seen through the lens of self-driving cars. But, most of the course focuses on topics we’ve never covered before, specific to computer vision techniques used in autonomous vehicles. There are plenty of new, valuable skills to be learned here!
Environment Setup and Installation
Get everything you need for the course installed: The Anaconda scientific Python development environment, the OpenCV computer vision package, the Tensorflow package for building artificial neural networks, and the code and data that make up the course materials.
Let's try out your new environment by doing real-time edge detection on a live video stream from your webcam, within a Jupyter notebook. We'll also do a quick overview of how Jupyter notebooks work.
Introduction to Self-Driving Cars
We'll cover the history of self-driving cars, which starts in 1925 and includes a lot of exciting progress that's been largely forgotten!
We'll quickly review the course outline, and give some guidance on which sections you might be able so skip given your prior experience.
Python Crash Course [Optional]
We'll briefly cover why whitespace is important in Python and how it's used, how to import packages of existing code libraries, and how to use lists in Python.
We'll continue diving into Python data structures with tuples and dictionaries, and examples of using them.
We'll cover the syntax of functions in Python, how to pass functions around as parameters, and lambda functions. We'll also see how boolean operations like equality and or work.
We'll see how for and while loops with in Python, and challenge you to a very simple exercise to practice what you've learned so far.
We'll walk through some examples of using the Pandas package, to slice and dice some fake video index data from cars.
We'll cover the different charts that MatPlotLib can produce from our data, and how to load and view images.
Seaborn both sits on top of Matplotlib to make it better, and introduces new kinds of visualization tools that can help you extract meaning from data. We'll walk through a bunch of examples using real fuel efficiency data for 2019 cars.
Computer Vision Basics: Part 1
Computer Vision Basics: Part 2
Computer Vision Basics: Part 3
Machine Learning: Part 1
Let's discuss how machine learning works, and how it fits in with the world of AI and deep learning.
Learn how train/test and K-fold cross-validation helps us to prevent "overfitting" models to the data they were trained with.
We'll go in depth on how linear regression learns how to fit a line to observed data, to create a simple model we can use to predict new observations.
Let's use linear regression to build a model mapping road conditions to vehicle speed.
Logistic regression builds a model that classifies data into one of two categories.
Let's practice using logistic regression to predict whether a car should go fast or slow, given distance to an upcoming bump and its size.
Decision trees build up a flowchart-like model, that classifies data based on various decision points that branch off to others.
In this activity, we'll implement the example from the previous lecture of predicting hiring decisions based on candidate attributes, and also see what happens when we use decision trees for the same logistic regression sample of predicting vehicle speed for an upcoming bump in the road.
Machine Learning: Part 2
We'll cover Bayes Theorem and how it can help us understand conditional probabilities, and apply it to Naive Bayes to classify email as spam or "ham."
We'll build a real spam classifier using Naive Bayes, and see how it well it works on our problem of classifying vehicle speeds based on upcoming obstacles in the road.
Support Vector Machines use the "Kernel Trick" to classify data. Hyperparameter tuning becomes important to find the right kernel to use, and the right parameters for that kernel.
We'll apply SVC to our vehicle speed classification problem, and illustrate hyperparameter tuning to find the best kernel and best set of parameters to use.
Artificial Neural Networks
Deep Learning and Tensorflow: Part 1
We'll talk about what Deep Learning is, and how Tensorflow works at a low level.
We'll explore how to construct deep neural networks for binary and multi-class classification with Keras, the importance of normalizing your input data, and how one-hot encoding is used to translate categories into a representation that's compatible with neural nets.
We'll use Keras to easily experiment with a variety of network topologies to apply deep learning to our "car approaching a bump" classification problem.
We'll go into more depth on activation functions and why ReLU is popular, and cover techniques for preventing overfitting including Dropout layers.
We'll run our previous neural network longer to make some overfitting happen, and see how a Dropout layer in Keras can improve accuracy by preventing overfitting.
Deep Learning and Tensorflow: Part 2
We'll introduce the concepts of CNN's, and how they are inspired by the biology of your visual cortex.
See how easy it is to implement a CNN using Tensorflow's Keras API.
Let's install tensorflow-gpu to make our CNN's run faster, and prepare our image data for training our CNN.
We'll classify real images from the cifar10 data set using our simple CNN, and analyze the results.
Max Pooling reduces image data down, which can dramatically speed up the processing of your CNN.
We'll improve on our image classification results by adding more layers to our CNN, and applying max pooling so we can run more epochs in a reasonable amount of time.