Page 79 - Data Science Algorithms in a Week

Machine Learning Applied to Autonomous Vehicles 63

selected based on intrinsic and extrinsic risk indicators. With the layers of planning and
control already set, a method is proposed to estimate the trajectory desired by the driver
during the cooperative control, allowing a decision to be made based on risk analysis.
Finally, different tests on VILMA01 (in the actual vehicle) are performed to validate the
proposed architecture.
These layers are not exactly a hierarchical model. Each layer has interactions at
different levels from directive to cooperative control with the others. These interactions
can be adapted depending on what the vehicle tries to do. For example, the architecture of
VILMA01 (Bedoya, 2016) aims to test strategies to drive a vehicle cooperatively
between an autonomous system and a driver which could help to reduce the risk of
accidents. This strategy assumes that the autonomous system is more reliable than the
driver, even though in other circumstances the driver could interact with the human
machine interface to disengage the autonomous system. Based on the architecture of
autonomous mobile robots, the proposed strategy is denominated as cooperative planning
and cooperative control, which determines when and how the driver can change the path
projected by the autonomous system safely through the steering. Figure 4 shows the
function blocks for the autonomous vehicle VILMA01. There are two important
considerations in the cooperative strategies. The first one is the interaction of the driver
and the robot through the steering (dotted line 1), which in turn generates the second one,
which poses the question in the planning layer (dotted line 2): is it safe to change the
projected path? These additions to the existent architecture generate two types of
cooperation. The first one, cooperative control is defined when the control signal of the
driver and the autonomous system cooperate during the local path planned by the
autonomous system. The second one (cooperative planning) is defined when the driver
and the autonomous system cooperate to change the local path after risk analysis is
performed.
Finally, the design of the layers, their functionality, and interactions can provide an
architecture its level of automation. According to Thrun et al. (2006), the six major
functional groups are interface sensors, perception, control, planning, vehicle interface
and user interface. Therefore, this layered architecture must take into consideration
hardware, software, and drive-by-wire automation.

MACHINE LEARNING APPLIED TO PERCEPTION

The most common applications of deep learning in autonomous vehicles are in
perception. As explained in the last section, one of the biggest problems in perception is
identifying objects on images because of the number of inputs which makes the
generation of a generic geometrical model very difficult. Therefore, it is a good problem
for deep learning.