Page 67 - Data Science Algorithms in a Week

Machine Learning Applied to Autonomous Vehicles 51

approach is to use machine learning to perform the perception function of the
environment; joint with complex mathematical algorithms to control the different
driver functions.
 Limited Self-Driving Automation (Level 3): The driver is sometimes in the
control loop and the vehicle is operating at level 3 automation. The most
common strategy for transferring control to the driver, particularly in high-risk
situations, is to use an emergency button. However, in practice, this may have
serious drawbacks. This issue is dealt with in a Google patent by Cullinane et al.
(2014), which describes a system in which all the security variables are checked
before the control is transferred. Tesla and the other actual autonomous vehicles
can be cataloged in this level of automation.
 Full Self-Driving Automation (Level 4): In this level, the driver is not expected
to take the control at any time during the desired trip. This level is fully
automated except for some environmental conditions. The actual systems without
the use of deep learning are far from being able to accomplish all the
requirements.

One of the research problems addressed by Autonomous vehicle is the lack of
driver’s attention (Kaplan et al., 2015). Several potential schemes have been introduced
by several researchers. The most important examples are:

 Jain et al. (2015) used a hidden Markov autoregressive input-output model to
capture contextual information and driver maneuvers a few seconds before they
occur, in order to prevent accidents.
 Malik et al. (2015) described an intelligent driver training system that analyzes
crash risks for a given driving situation. This opens possibilities for improving
and personalizing driver training programs.
 Liu et al. (2014) proposed a method for predicting the trajectory of a lane-
changing vehicle using a hidden Markov model to estimate and classify the
driver’s behavior.
 Amsalu et al. (2015) introduced a method for estimating a driver’s intention
during each step using a multi-class support vector machine. Although the
approaches described in these studies yield satisfactory results, none of them
specifically handle cooperative control between automated intelligent systems
and the driver.
 Merat et al. (2014) described tests in a simulator to investigate driver’s behavior
when the driver is resuming manual control of a vehicle operating at a high level
of automation. Their study sought to contribute to an understanding of suitable
criteria for the design of human-machine interfaces for use in automated driving