Page 77 - Data Science Algorithms in a Week

Machine Learning Applied to Autonomous Vehicles 61

"Dynamic World Modeling" is the problem by which an internal description of the
environment is assembled using proprioceptive sensors. By dynamic, it is meant that the
description evolves over time based on information from perception. This description is a
model because it permits the agent to represent the external environment. Fusion
techniques have been used to combine the measures provided by the sensors and their
comparison with the respective mathematical models of the robot and the environment.
Perception and state estimation have many characteristics in common. State
estimation calculates the state of the vehicle. On the other hand, perception estimates the
state of the environment. Although state estimation tends to deal with signal variations
over time, perception tends to deal with signal variations over space. In this layer,
machine learning techniques have been used because the proprioceptive sensors generate
vast amounts of information. This information has to be processed in a timeless fashion
and therefore conventional techniques are not able to handle this online. For example, the
amount of information generated by a camera is very high: If you have a color camera in
full HD, it generates more than six million of points (two million pixels by each of the
three basic colors) at a rate of 30 frames per second. This information must be processed
in real time in order to obtain the characteristics of the environment like traffic signals,
pedestrians, cars, and bicycles.

Figure 3. Layers in the mobile robotics architecture (Bedoya, 2016).

The planning or navigation layer will determine where the vehicle should go
according to the perception and the mission. This has to include a risk analysis to
determine the path and speed of the vehicle. The cognition aspects of an autonomous
vehicle depend on the mobility capabilities which are studied by the robotics navigation