1DATA SYNCHRONIZATION STRATEGIES FOR
MULTI-SENSOR FUSION
Nico Kaempchen, Klaus Dietmayer
University of Ulm, Department of
Measurement, Control and Microtechnology
Albert-Einstein-Allee 41, D-89081 Ulm, Germany
Tel: +49 731 50 26326 - Fax: +49 731 50 26301
E-mail: {Nico.Kaempchen, Klaus.Dietmayer}@e-technik.uni-ulm.de
SUMMARY
Advanced driver assistance systems (ADAS) have increasing demand for several sensor
systems. Fusing multi-sensor data enlarges the field of view and increases the certainty and
precision of the estimates. A crucial part of a fusion system is the data association, which
requires data synchronization. The major synchronization strategies for data fusion are
discussed and contrasted with respect to their usability in ADAS.
INTRODUCTION
Recent projects on driver assistance systems have focused on applications such as Pre-Crash
(Chameleon) [1], ACC Stop&Go (Carsense) [2] and recognition of vulnerable road users
(PROTECTOR). These advanced driver assistant systems (ADAS) have increasing demand
for several sensor systems, which are complementary but also redundant. Much research has
therefore been focused on high-level multi-sensor fusion [2, 3, 4, 5]. The aim of such a step is
to provide a fused description of the traffic scene surrounding the vehicle, which is relevant
for ADAS, but not specific for a certain application. This fusion system incorporates the data
of the diverse sensors into a single description. Thus the field of view of a single sensor is
enlarged, the certainty and precision of the estimates is increased and additionally this system
design is economically efficient, as different applications share a set of sensors.
A crucial part of such fusion systems is the data association. In order to update the object
states of the environment description, the sensor data must be associated with the environment
description. Correct association requires a synchronization of the sensor data and the
associated object state. Different synchronization strategies will be discussed with respect to
the common automotive sensors and applications.
SENSOR FUSION ARCHITECTURE
Established advanced driver assistant systems, such as lane departure warning (LDW) and
automatic cruise control (ACC) have a common architecture, as shown in Figure 1a. The
application is implemented for a special sensor configuration and therefore depends on these
sensors. With an increasing number of applications which share a set of sensors, the sensor

2dependent part of the software would exist several times, i.e. found in each application (see
Figure 1b). Thus emerges a strong motivation to separate the sensor specific part of the
system and use it as an interface between the sensors and the applications (see Figure 1c). The
sensor fusion module is responsible for collecting measurements, interpreting them sensor
specifically and incorporating them into a unified, consistent description which is then
forwarded to the applications. The sensor fusion must be in part sensor specific in order to
obtain a maximal profit of each sensors measurement.
Sensor Fusion
Application
Sensors
a b c
Figure 1: Driver assistant system architectures. (a) A single application which depends on
one or several sensors. (b) Several applications each based directly on one or several
sensors. (c) Possible future advanced driver assistant system architecture with common
sensor fusion layer, thus separating the application from sensor specific implementations.
A general architecture of a sensor fusion system is contains the basic components shown in
Figure 2 [6]. The environment description is composed of a set of objects, each of which is
defined by an object state. There are different object models for trucks/busses, passenger cars,
(motor) bikes, pedestrians and for stationary objects. Additionally the road is modeled as well
as the ego-vehicle.
The objects of the environment description are predicted in order to synchronize them with
the incoming sensor data. The object states are transformed into the parameter space of the
sensor data by the use of inverse sensor models and the association is performed. In case of an
established association the sensor data is integrated into the object state. An object
management handles the generation and deletion of objects in the environment model.
As mentioned, the sensor fusion is not independent from the sensor’s characteristics. The
sensor specific modules are the inverse sensor model and the association. These are the only
components of the sensor fusion layer which need to have some knowledge of the sensor’s
characteristics. The integration and prediction may be performed using a Kalman-filter.