Car Factory
  • Sukhjeet Singh

Deep Learning in Autonomous Vehicles

Ever wonder how the google language converts an entire webpage into an entirely different language in a matter of seconds and how your phone groups images in your gallery based on their location? All of these are a product of deep learning. Deep learning is a subset of machine learning which in turn is a subset of artificial intelligence. Artificial Intelligence is a technique to enable machines to mimic human behavior, machine learning is a technique to achieve artificial intelligence through algorithms trained with data, and finally, deep learning is a type of machine learning inspired by the structure of human brains. In terms of deep learning, the said structure is called an artificial neural network. Deep learning distinguishes itself from machine learning by the type of data that it works with and the methods in which it learns. While machine learning typically predicts the outcome based on a trained data set, deep learning using artificial neural networks attempts to mimic the behavior of the human brain through a combination of data inputs, weights, and a bias. These elements work together to accurately recognize, classify, and describe objects within the data. Unlike classical machine learning, artificial neural networks also use intermediate layers for data optimization, making it more reliable, efficient, and accurate than classical machine learning in predicting and reorganizing outcomes of data. Having these advantages, deep learning is used in multiple fields of technology including but not limited to computer vision, customer experience, language recognition, and autonomous vehicles.

Experiments on making an autonomous car have been performed since 1920 and some promising trials took place in the 1950s. The first self-sufficient and a fully autonomous car appeared in the 1980s with Carnegie Mellon University's Navlab and ALV projects in 1984, Mercedes-Benz and Bundeswehr University Munich's Eureka Prometheus Project in 1987. With these, major companies such as Mercedes-Benz, General Motors, Continental Automotive Systems, Autoliv Inc, Bosch, Nissan, Toyota, Audi, Volvo, Vislab from the University of Parma, Oxford University, and Google also started investing in autonomous vehicles. With driver-less cars and deep learning implementation in autonomous cars with all other electronic components of advanced technology, it surely provides a soothing feeling to the public, but as every coin has two sides. There are both advantages and disadvantages to these autonomous cars.

Making automatic systems as complex as autonomous vehicles is a real challenge as one has to deal with processing loads of data at a given point in time. But if we have to define where deep learning is used in autonomous vehicles, then there come four basic yet very effective areas where deep learning is applied and helps a vehicle to localize from point A to B.

1. Perception

Perception is used by the system in autonomous vehicles to find the environment and obstacles around the vehicle. It is one of the pillars on which the vehicle relies and contains deep learning.

For perceiving the lanes and obstacles around you, the autonomous vehicle uses three sensors i.e. camera, LiDAR (Light Detection and Ranging), and RADAR (Radio Detection and Ranging). These sensors are used in combination or independently to detect the lanes and surroundings of autonomous vehicles. Algorithms such as LaneNet are used for detecting lanes, and YOLO or SSD is very popular among designers of neural networks in deep learning for autonomous vehicles.

2. Localization

Localization is one of the second pillars and building systems for autonomous vehicles. Localization is all about the location of the vehicle in the world with the highest possible accuracy. Conventional location determination technologies like GPS are not at par with the accuracy of localization methods used in autonomous vehicles and developed via deep learning.

Based on the type of algorithm used for localization, we have mainly 3 methods discussed as follows:

  • Knowing the Map and Initial location - It is when the vehicle knows the initial location and the destination and has access to the map so it just needs to follow the route to reach the destination.