Signal Processing - March 2017 - 61

than positioning, since a true longitude and latitude position is
of no value unless the map and situational awareness have the
same absolute accuracy.
Consider the schematic picture of a vehicle in Figure 1.
The trend is to make vehicles autonomous [1]-[4] and utilize
advanced driver assistance systems (adas). Hence, there is a
need to improve both localization and velocity estimation systems. Basically, it is going beyond traditional point estimation
methods [5], [6] to get a better probabilistic understanding [7]-
[10] of the environment using more detailed models and filters. The actuators (brake, steering wheel, engine torque) have
essentially been the same since the automobile was invented,
and only a few new actuator concepts have been introduced
(active suspension, movable headlights, etc.).
In stark contrast to the actuators, the number of sensors has
increased substantially over the last decade [11]-[13], e.g.,
■ the inertial measurement unit (Imu) [14] in the engine
control unit (eCu) and in suspension sensors for estimating the vehicle state
■ vision, stereo vision, night vision, radar, sonar for monitoring the surroundings, and keeping the vehicle in the lane at
a safe distance (i.e., relative position control)
■ the wheel speed sensors (Wss) introduced with the antilock braking system (aBs) are one of the most versatile
sensors in the car
■ databases such as vectorized road maps [15]-[18] utilized
for positioning including road height, map matching [11],
[19]-[22], and pothole indications [23], etc.
Cars are slowly following the development of smartphones.
Today there are many radio receivers in vehicles: cellular network, Bluetooth, and Wi-Fi. These can be used in various
signal processing applications such as localization and speed
estimation. It is less explored that these information sources
all include indirect information about the position. The vehicle
state sensors contain information of road signatures (curves,
banking, slopes, and small variations in the surface height).
The vision sensors can see landmarks of known position. How
the Wss can be used for positioning is described next.
sensor fusion is used in all of the aforementioned cases for
refining the information, where there are several good examples of virtual (or soft) sensors that compute physical quantities that cannot readily be measured by sensors. examples
include the detection of obstacles, pedestrians, and animals
from vision sensors and tire pressure and road friction from
Wss. Our approach is based on statistical signal processing
techniques, based on a simple odometric model of the vehicle
and a model of each sensor relating to the vehicle state. In particular, the road map information is nonlinear and cannot be
approximated with a linear Gaussian model, so a PF framework is preferred to Kalman filter (KF) algorithms. The sensor
fusion concept is summarized in Figure 2.

Future localization algorithm applications
This section discusses the need for improved localization
algorithms highlighting areas such as cloud-based computations, autonomous driving, handheld devices, and mapping.

Driver
and
Planner

Controller

Actuators
Vehicle
Sensors

Sensor
Fusion

Figure 1. An illustration of data flow in a vehicle. Future AdAs functionality might include cloud information as well as control, sensor fusion, and
planning.

Cloud-based services
To some extent, positioning today is used for cloud-based
crowdsourcing, such as in apps for pothole detection and
speed camera positions, among others. This is an area that
probably will explode in the future when manufacturers integrate these reports in their own servers and offer their own
and other customers services based on this information.
Consider potholes as an example of a virtual sensor: they
are annoying to passengers and may be a hazard to the vehicle. These are easily detected by accelerometers, Wss, or suspension sensors, and the presence of potholes can be included
in the car's navigation system. The problem is how to share
the information between users. Figure 3 shows an illustration of pothole detection and clustering [23]. many vehicles
have, in this case, hit the same potholes and delivered the
estimated position to a cloud database. a cloud-based clustering algorithm is then used to merge the various pothole
detections into one unique pothole, and possibly also project
the position to the road. This information can now be shared
with other drivers, but it could also be used by road authorities for maintenance.

Autonomy
Future autonomy will put high demands on the localization
algorithms. despite the media success of self-driving cars, the
technology is still in development. On one hand, there is the
defense advanced Research Projects agency (daRPa) generation of cars where localization is based on a laser scanner,
however, the cost of these vehicles is still far from affordable;
see the daRPa grand challenge and urban challenge [1], [2].
Further, the laser scanner's raised placement on the rooftop is
not well aligned with the design. Google's self-driving car [3]
is equipped with laser, radar, and cameras on the rooftop.
apart from most vehicles, it is not designed to be, or even
possible to, drive manually. On the other hand, we have seen
self-driving cars (audi Rs7 Piloted driving presented in
Hockenheim, Germany) positioned using differential GPs,
including yaw estimation from multiple antennas and camera
information. These cars have demonstrated spectacular performance on restricted accurately mapped areas.
autonomous functions in the car, and in the extreme selfdriving cars, will need another level of integrity. The localization
algorithms must work in tunnels, parking garages, urban street

IEEE SIgnal ProcESSIng MagazInE

March 2017

Table of Contents for the Digital Edition of Signal Processing - March 2017