AI LITERACY - A Podcast about Artificial Intelligence
#8 How a self-driving car learns how do drive with Saquib Sarfraz (Mercedes Benz-Daimler)
Mon Jan 31 2022Description
Computer vision and deep learning are revolutionizing the development of self-driving cars. This episode explores the role of computer vision in enabling autonomous vehicles, the training process using data and deep learning, the workings of neural networks and prediction, behavior intent prediction and risks associated with self-driving cars, and the road to fully autonomous driving. The journey towards fully autonomous driving will be gradual and require addressing various factors such as infrastructure and public trust.
Insights
Computer vision and deep learning are key technologies for self-driving cars
Computer vision technology, accelerated by deep learning, plays a crucial role in enabling self-driving cars. Neural networks are trained with vast amounts of data to recognize objects in their surroundings.
Behavior intent prediction is an active area of research
Machine learning techniques, particularly neural networks, are used to predict behavior on the streets. While not yet 100% reliable, intent prediction can signal control back to the driver in most cases.
Risks and challenges slow down the adoption of self-driving cars
The complexity and risks associated with self-driving cars, such as software shutdowns and potential accidents, have caused many companies to postpone or revise their plans. Factors like insurance coverage and road infrastructure also contribute to the slower rollout of autonomous vehicles.
Fully autonomous driving will take time and require significant changes
Achieving fully autonomous driving on all roads will be a gradual process, similar to the transition from horses to vehicles. While level 4 autonomous driving may be seen in the next five years for specific applications, level 5 autonomy without human intervention will take longer than anticipated.
Widespread adoption of higher levels of automation depends on regulatory approval and consumer trust
To achieve widespread adoption of higher levels of automation, such as level 4 and level 5 autonomous driving, regulatory bodies need to provide approval and consumers need to trust the technology. These factors are crucial for ensuring safety and reliability.
Chapters
- Introduction to Computer Vision and Self-Driving Cars
- Training Autonomous Vehicles with Data and Deep Learning
- Understanding Neural Networks and Prediction
- Behavior Intent Prediction and Risks of Self-Driving Cars
- The Road to Fully Autonomous Driving
Introduction to Computer Vision and Self-Driving Cars
00:05 - 09:10
- Dr. Sakeev Sarfas, a senior scientist at Mercedes-Benz Daimler, discusses computer vision and its role in enabling self-driving cars.
- Computer vision was already an active area of research before the popularity of deep learning, which has accelerated performance in tasks related to AI.
- The availability of larger compute resources and access to more data has allowed for advancements in computer vision technology.
- Autonomous vehicles are just one exciting application of computer vision technology, with potential for a huge impact on society.
- In a couple of decades, it is likely that people will look back and wonder how we ever let humans drive when autonomous vehicles become the norm.
Training Autonomous Vehicles with Data and Deep Learning
08:42 - 17:14
- To achieve self-driving cars, they need to mimic human behavior by mapping the surroundings, perceiving objects, predicting actions, and planning reactions.
- The main types of data used to train autonomous vehicles are camera images and LIDAR data.
- Autonomous driving technology has been developed for almost two decades, with real-world driving data collected over time.
- Deep learning and neural networks are essential technologies for training computers to recognize objects in their surroundings.
- Deep learning involves training neural networks with vast amounts of examples to learn object recognition in videos and images.
Understanding Neural Networks and Prediction
16:48 - 24:32
- Neural networks simulate the biological neurons in our brain through layers of artificial neurons.
- Artificial neurons are simple functions that multiply inputs with numbers and add them to produce an output.
- Neural networks learn to map inputs to outputs by learning these numbers through examples.
- Neural networks have multiple layers or levels of filters that learn different features in a hierarchy.
- The prediction part of the process involves classifying intent and predicting future events based on the learned environment.
Behavior Intent Prediction and Risks of Self-Driving Cars
24:02 - 32:16
- Machine learning can be used to predict behavior on the streets, such as when a person is going to cross the street or catch a bus.
- Neural networks can accurately recognize pedestrians and their poses, allowing for intent prediction.
- Intent prediction is not yet 100% reliable, but it can signal control back to the driver in most cases.
- There is still a trust issue with self-driving cars and their ability to predict what will happen on the roads.
- The complexity and risks of self-driving cars have caused many companies to postpone or take back their plans for autonomous vehicles.
- Risks include potential software shutdowns and damage caused by self-driving cars hitting something.
- Factors such as insurance coverage and road infrastructure also slow down the rollout of self-driving cars.
- Self-driving technology is being introduced progressively through different levels of automation, from basic driver assistance to full autonomy without human intervention (level five).
- Most current autonomous driving vehicles operate at level two or three, with level four being more advanced but not yet widely available.
- Approval from regulatory bodies and consumer trust are necessary for widespread adoption of higher levels of automation.
The Road to Fully Autonomous Driving
31:47 - 35:44
- Fully autonomous driving is currently limited to predefined areas with no other vehicles around.
- The switch to level 4 autonomous driving is not expected in the next five years due to various factors.
- Progress towards fully autonomous driving will happen gradually, similar to the transition from horses to vehicles.
- Level 4 autonomous driving may be seen in the next five years for taxi services or similar applications.
- Achieving level 5 autonomous driving, where there is no human intervention, will take more time than anticipated.
- A McKinsey study predicts that by 2040, up to 66% of passenger kilometers driven may be accounted for by autonomous cars.
- However, achieving fully autonomous driving on all roads will require significant changes and may take around 15-20 years or more.