Signal Processing - November 2016 - 81

by one system are also detectable by the other systems. In
addition, for the detailed sensors typically used for autonomous vehicles, communicating raw sensor observations is
probably not feasible; thus, compression and semantic labeling is needed.

Performance of the OCP in the presence of
communication and sensing impairments
To illustrate the role sensing and communication play in solving the coordination problem, we consider an intersection
scenario of the type illustrated in Figure 2, where incoming
vehicles periodically measure and send their state information
(UL) to a centralized controller. The control is performed in a
receding horizon fashion, where the controller solves a finite
time OCP, and broadcasts the resulting control actions to the
vehicles (DL). We simplify the OCP by modeling vehicles
as points with positions x i (t) , velocities xo i (t) , and controls/
accelerations u i (t) = xp i (t) along one-dimensional trajectories,
aligned with the center of each road. The intersection is then
modeled as an interval [L i, H i] on each trajectory. The objective (4a) is chosen to be
tf

J i (x i (t), u i (t)) = Q i

(v ref
i

- xo i (t)) 2 dt + R i

u 2i (t) dt, (5)

where v ref
i is a constant reference speed, t f is a time horizon,
and Q i 2 0, R i 2 0 are weights set by the user. The liveness
constraint (4f) is stated as x i (t f ) $ H i for all vehicles. Finally,
the problem is discretized and solved using standard optimization tools.

Ve
h

icl

Ve
h

e3
icl

Ve
h

icl

Time
Slot 1

15
Position (m)

part needs to be solved online [28]. However, offline mapping is time-consuming and may need to be repeated periodically. In contrast, under SLAM, there is no need for
offline mapping, rendering it less sensitive to changes in the
environment. However, the SLAM problem is inherently
more difficult than self-localization in a preconstructed
map and thus tends to give inferior positioning accuracy.
The problem of estimating the position of relevant road
users, including uncertainty measures, is known as a multisensor and multiobject tracking problem, which is a wellstudied problem within several applications. In contrast to
the classical formulation, objects in an automotive setting
typically give rise to multiple radar and lidar measurements,
thus violating the classical point-source assumption (one
measurement per object). Instead, objects such as vehicles,
need to be treated as extended objects, which is less studied
and typically leads to more complex algorithms. However,
including multiple measurements per object also allows for
a richer description of the object such as orientation and
physical dimensions.
Both self-localization and estimating the position of
other road users can be performed cooperatively [29]. For
instance, for the latter problem, in addition to exchanging
position estimates, information about the physical extension
can be shared, thus greatly simplifying the inference and
reducing the uncertainty in the position of the objects. However, as the sensor observations are typically not labeled, to
use the measurements properly we need to be able to correctly associate them with the information coming from
the other vehicles and accurately match them to the local
view of the traffic situation, adjusting for delays due to
data transmission and asynchronous sensor operation. For
self-localization with offline mapping, the map resides in
the cloud and can thus easily be shared among the cooperating vehicles. By sharing position estimates in the joint
map, together with uncertainty measures, each vehicle can
jointly estimate a more accurate ego-position as well as the
position of all the other vehicles by fusing with the local
perception from the on-board sensors [30]. This way, the
self-location problem and positioning of other road users are
solved simultaneously. This also leads to the possibility of
quickly detecting and sharing changes in the map (e.g., the
construction site in Figure 6). To increase the positioning
accuracy, estimates of relative position to a selected set of
high-quality landmarks can be exchanged between the vehicles and used in a similar manner. For SLAM, cooperation is
also beneficial. There are two types of cooperative-SLAM
(C-SLAM): centralized and distributed. In the former, the
cooperating systems communicate their position estimates
and current sensor observations to the cloud where a joint
map is formed and shared among the systems [31]. In the
distributed versions, however, this information is instead
communicated to the individual vehicles, which build and
keep their own map using all the information. Both of these
C-SLAM methods require that the cooperating entities have
a fairly homogeneous sensor setup such that landmarks seen

10
5
0

Time
Slot 6

−5
−10
−15
−20
11.5 12

12.5 13 13.5 14 14.5
Time (s)

15.5 16

Figure 7. The visualization of a part of the position trajectories along
each road for six coordinated vehicles under perfect communication and
sensing. For each vehicle, the intersection starts at 0 [m] and ends at
10 [m] . The colored lines represent the trajectories of each vehicle. The
correspondingly colored boxes visualizes the time slots during which the
intersection is occupied by each vehicle. Note that collisions would occur
if the time slots were to overlap. In this idealized case, the time slots are
tightly packed. Hence, there is no safety margin, and the performance of
the system in terms of the objective (5) is pushed to its limits.

IEEE SIgnal ProcESSIng MagazInE

November 2016

Table of Contents for the Digital Edition of Signal Processing - November 2016