Você está na página 1de 6

A Modular System Architecture for Sensor Data

Processing of ADAS Applications


Michael Darms, Hermann Winner

Propulsion & Drivetrain


Abstract— In this article a modular system architecture for Secondary Safety Systems
fusion of data from environment sensors for advanced driver- Vehicle Stabilising (Primary Safety)
assistance systems (ADAS) is proposed. The architecture allows Vehicle Conducting Systems
different applications to have access to the fused sensor data by Information Systems
processing the data with respect to specific demands of different

Drive train management


Vehicle Stability Control

Transmission Control/
application groups. In the article the growing interconnection of

Engine Management
Pre-Crash Restraint
Suspension Control
functions is illustrated by two examples and the most relevant Applications

Navigation
consequences of this trend are given. The question, if an

ACC
application independent processing of sensor data is feasible is
discussed and the most important aspects for the design of the
system architecture are given. This provides the basis for the
explanation of the system architecture. Components
Surround Sensors
Long Range Radar X
Index Terms— ADAS, Adaptive Filter, Sensor Fusion, System GPS + Digital Map X (x)
Architecture Vehicle Dynamics Sensors
Wheel Speed X X X X X (x) X
Longitudinal Acceleration X (x) (x) X (x)
Lateral Acceleration X (x) (x) (x)
I. INTRODUCTION Yaw Rate X X X X X (x)
Steering (Wheel) Angle (x) (x) X X (x)

A DVANCED Driver Assistance Systems (ADAS) rely on


information about the vehicle environment provided by
different types of sensors, like Radar, Video or Lidar for
Actuators
Brake Torque
Engine Torque
X
X
X
X X X
Restraint actuators X
example. Up to now the information provided by a sensor is Wheel Spring/Damper (x) X
usually connected to a specific application directly, like Radar
Fig. 1. Allocation of Sensors and Actuators to Applications (Today).
to Adaptive Cruise Control (ACC) or Video to Lane
Departure Warning (LDW). In the future costs will be reduced
disadvantages.
by making the information delivered by (environment) sensors
In the article the question, if an application independent
available for use by different applications (see for example
data processing is practicable is discussed. Then the main
[1], [2]). In the first part of the article the growing
aspects for the development of a system architecture due to
interconnection of functions is illustrated and the main
the multisensor approach are explained. The architecture itself
consequences of this trend are pointed out.
is then explained in section IV with reference to the results of
One of the most important aspects is the structure
the preceding sections.
(architecture) of data processing in terms of software and
hardware. Theoretical models (see for example [3], [14]) have
II. GROWING INTERCONNECTION OF FUNCTIONS
been developed for different data fusion architectures, each of
which is associated with specific advantages and A. Brief look into History
Electronic linking of systems in the automobile started in
the Eighties with the development of the electronic
transmission control and the traction control system, which
Dipl.-Wirtsch.-Ing. Michael Darms is research associate at the Department
of Automotive Engineering at Darmstadt University of Technology, both interact with the engine management. In the first
Petersenstr. 30, D-64293 Darmstadt, Germany (e-mail: mdarms@fzd.tu- generation of these systems, it was only possible to use
darmstadt.de). dedicated point to point connections for signal transfer
Prof. Dr. rer. nat. Hermann Winner is head of this department (email:
hwinner@fzd.tu-darmstadt.de). (mostly PWM based). Specific functions were involved such
This article was prepared within the framework of a joint research project as, for example, ignition advance angle adjustment for rapid
with Continental Teves AG & Co oHG, Guerickestr. 7, D-60488 Frankfurt, engine torque reduction. For a longer lasting torque reduction
Germany. Dipl.-Ing. Maxim Arbitman is the project manager.
of the traction control system, even a stand-alone component
Propulsion & Drivetrain
Secondary Safety Systems
Vehicle Stabilising (Primary Safety)
Vehicle Conducting Systems
Information Systems

Applications

Side Obstacle Detection/

Lane Departure Warning

Drive train management


Vehicle Stability Control
Lane Change Assistant

Lane Keeping Support


Adaptive Light Control
Park Distance Control

Transmission Control/
(Semi-)Active Parking
Congestion Assistant

Engine Management
Vision Enhancement

Pre-Crash Restraint
Suspension Control

Collision Mitigation
ACC + Stop&Go
Backing Aid
Navigation
Components
Surround Sensors
Short Range Sensor (front) X X X X X (x)
Short Range Sensor (side) (x) X X (x)
Short Range Sensor (rear) X X (x)
Long Range Radar (x) (x) X (x) (x) (x) X
Mono-Video (x) (x) X X (x) (x) (x) (x) (x)
Stereo-Video (x) (x) X X X X X (x) X
GPS + Digital Map X (x) (x) (x) (x) (x) (x) (x)
Vehicle Dynamics Sensors
Wheel Speed X (x) X (x) (x) X X X X X X X X X X (x) X
Longitudinal Acceleration (x) X X (x) (x) (x) (x) X (x)
Lateral Acceleration (x) (x) X (x) X X (x)
Yaw Rate X (x) (x) (x) (x) (x) X X (x) (x) X X X X (x)
Vehicle Travel X (x) X X X (x)
Steering (Wheel) Angle (x) (x) (x) (x) X X (x) X X X X X (x) X
Actuators
Brake Torque X X X X X
Engine Torque X X X (x) X X
Steering Angle/Torque X X (x) (x) X
Restraint actuators X (x)
Wheel Spring/Damper (x) X (x)

Fig. 2. Allocation of Sensors and Actuators to Applications (Future).

was developed with which the air feed was controlled in series expansion in networking.
with the throttle. If you analyze this matrix, you get 135 links. In relation to
Another reason for the networking of functions occurred as an application, about 8 components are used, which in turn
a result of the desire to save on redundant sensors (for must be available to roughly 7 applications. The savings
example: use of ABS rpm sensors for transmission control and gained by networking are significant (135-18=117); the same
tachometer). applies to the corresponding increase in interfaces.
As a result of the foreseeable electronic networking, several Even if this compilation of components and
approaches towards a vehicle compliant data bus system came sensors/actuators is more or less arbitrarily, it is possible to
about at virtually the same time from which CAN made its derive from this that approximately five times more functional
mark as a standard. interfaces are created as a result of the multiple use of
components.
B. Present Situation
If you examine current applications in a car, a matrix can be D. Consequences
created, which allocates sensors and actuators to applications The use of information from different sensors in multiple
(see Fig. 1; components, which only serve one core function, applications results in a high level of interconnection. The
are not listed). In this selection, on the one hand one of the most important consequences are:
eleven components on the average serves 3.5 of the eight • Bandwidth, network topology. Even today's data
applications. On the other hand every application uses about networks make full use of the entire bandwidth, which usually
5.5 commonly used components. In total, there are 39 links. means that several bus systems find their way into higher class
The number of components has been reduced by 39-11=28 cars. These bus systems are assigned to different clusters
due to networking. They are replaced by 28 system-wide which exchange very little data with one another by bridges or
interfaces. gateways (e.g. power train, body electronics, multimedia). A
higher degree of networking results, on the one hand, in a
blurring of the clearly defined limits of the clusters, and, on
C. Future Situation
the other, in an overall increase in data volume. Bus systems
If you look further into the future, a significantly higher such as FlexRay or TTP, which allow higher transmission
degree of networking will be attained in a higher value car rates than CAN, offer useful solutions.
(see Fig. 2; for further simplification, interconnections due to • Software redundancy. Sensor signals are further
X-by-wire systems, were ignored). Driver assistance and processed and combined with other signals using model
safety systems, which for their part are being introduced with assumptions. If each function processes the data for itself,
new environment sensors, are mainly responsible for the
fx Precision
fx A
fx
fx

t
B
t4
t3 Delay Time
t2 Fig. 4. Demands for Delay-Time and Precision from different Applications
t1 (A, B)
x
times with the corresponding probabilities, at least two models
Fig. 3. Four distance measurements (x) at four different times (tn) with may be correct: you may have observed an object moving
corresponding probabilities (fx). One possible model explaining the data is an
object moving with constant velocity (solid line); another possible model is a
with a constant velocity or an object which is standing still.
non moving object (dashed line). The measured data can be explained by both models, even if
the first one is more likely in this example.
significant overlapping occurs and similar software modules What you can also see from this example is that the
are developed without being coordinated with one another. model/assumption describing the dynamic behavior best can
• Reliability, failure behavior. Reducing the number of only be determined after the gathered data has proved it to be
components by interconnecting functions generally leads to an significant.
increase in the reliability of the system as a whole. However, Thus depending on the application, assumptions regarding
the failure of a given single component that is essential to the choice of a model must be made beforehand in order to
several systems can result in the failure of all of these systems. achieve the best estimation for the relevant situation “in time”.
• Development process. Despite the advantages offered The estimation may also be incorrect if the assumptions are
by computer-aided development support, it becomes wrong. Thus a trade-off between precision and delay time has
significantly more difficult to maintain a complete overview to be made, based on the function of the application (see also
of systems characterized by extensive interconnection of Fig. 4).
functions, even if comprehensive specifications are defined. In the case of an ACC system, the assumption that a vehicle
Furthermore, several different partners, some outside the moves as would be expected for a “normal driving” may be
company, are involved in the development of the overall appropriate in nearly all situations encountered on a highway.
system. Given a reasonable structure for the overall system The assumption is usually not applicable to safety systems,
and corresponding interfaces, the development process can be however. When filtering the data based on the “normal
efficiently arranged and further development can be controlled driving” assumption, the state estimation of an object may
more effectively. well be too “conservative” for a collision-avoidance algorithm
(see also [8], [10]).
III. SENSOR DATA PROCESSING IN VIEW OF DIFFERENT
APPLICATIONS B. Adaptive Filters
As a way to cope with the problem of varying models, one
A. Model Based Filtering usually uses an adaptive filter. A very common approach is
The use of the same sensor data for different applications as the Interacting Multiple Model Estimator (IMM, see e.g.
described above raises the idea, to have an application [11]). This approach assumes that the observed object is in
independent data processing. Using the raw data, one would one of a finite number of modes, which can be described by a
design an estimator that processes the data in such a way, that corresponding suitable model. The algorithm then does a soft
one has the best estimate that all applications can work with. switching between these modes according to computed mode
There is still the question, if this is feasible (see also [10] probabilities [11, p. 466].
for example). An estimated value is defined using a model of Nevertheless the models/modes are motivated by the
the dynamics of actual parameter and the knowledge of the application. If you do not expect a car doing a full brake, it is
measurement error. A very common estimator for dynamic not necessary to have a model for this mode available. It may
systems is the (Extended) Kalman Filter (KF, see e.g. [11]). even be a bad decision to have too many (and maybe similar)
The Problem is that there is no such thing as a "correct" models in the IMM estimator, as the algorithm then may not
model, especially when you observe an object over a longer be able to “distinguish” between them, causing a lack in filter
period of time. An observed car may stand still, it may move performance [11, p. 475].
with a constant velocity or a constant acceleration, for Furthermore adding a new application requiring a new
example. model to the overall system will modify the behavior of this
Fig. 3 explains this problem with a constructed example. If “system wide” filter and you would have to make sure again
you look at the four distance measurements at four different that the other applications can work with this modification
without a loss in performance and without errors. V. ARCHITECTURE ASPECTS
C. Application based Data Processing A. Centralized versus Decentralized Architecture
As a result of this discussion it seems to be more efficient to With regard to data fusion at object level, literature
have different data processing for different applications. As a distinguishes in general between three different configurations
way to reduce complexity one can define application groups, (see e.g. [14]): centralized, autonomous (decentralized) and
which rely on the same filter/estimator, even if this might not hybrid data fusion architectures. The configurations are
be optimal in all circumstances. (Remark: the use of the IMM distinguished in greater detail in [3]. This article discusses
estimator as such an application specific filter may still be a only the two extreme configurations (centralized and
good choice). decentralized) in order to illustrate the most important
differences.
IV. SENSOR FUSION ASPECTS In the centralized version, all data processing is performed
in a central unit. All available information (raw sensor data) is
A. General Aspects
consolidated in a central unit, without pre processing. A
Besides having the data from one sensor available for central database for all applications is created on the basis of
different applications it is also possible to combine the data an all-encompassing model.
from different sensors. The opposite pole is the decentralized structure. In this
In general terms, the following four fundamental aspects configuration, all data is processed by the sensor (model-
apply to a multisensor system (according to [6], [7]): based filtering, object tracking). The tracks generated in this
• Redundancy. Measurement precision can be enhanced way are fused at a higher level using a track-to-track
by using redundant information. In addition, the fault correlation based on the estimated states in the sensors.
tolerance of the overall system increases, as the failure of one Both configurations offer advantages and disadvantages. In
sensor does not necessarily result in the failure of the system the central configuration, all data is transmitted to the central
as a whole. unit without loss of information and can be processed with
• Complementarity. Means that information is enhanced maximum accuracy. Furthermore, there is no danger of
by input from different sensor types. This can be visualized inconsistencies due to dissimilar model assumptions in the
both spatially (e.g. expanding the covered range) and in different processing units.
relation to an observed object (e.g. independent position and On the other hand, a large bandwidth is required to transmit
velocity measurements). Where applicable, ambiguities can data from the sensor to the central unit. All processing power
also be resolved using complementary sensors. is concentrated at one point in the architecture and has to be
• Time precision. Acquisition speed can also be increased dimensioned accordingly. The most significant disadvantage
with the aid of multiple sensors. This can be accomplished, for is that it will be difficult to modify or extent the system in any
example, through parallel processing of information or even way, as every change requires reoptimization of the central
through an appropriate chronological configuration of sensor unit.
information. In the second case, (decentralized architecture) only a small
• Overall costs. Costs and benefits associated with the use bandwidth is needed to transmit the information (tracks in this
of multiple sensors must be compared with the costs and case). System modifications/expansions can also be realized
benefits of a system with fewer sensors. with relative ease, as all data is processed in the sensor.
B. Data Association One problem associated with the use of the decentralized
If there is more than one object to be tracked (“multitarget arrangement is that the consistency of model assumptions over
tracking”), measurements acquired by the sensor(s) must be the whole system is not guaranteed. Different sensors may use
associated to the specified tracks (“data association”; see [3], different assumptions to filter the data. Furthermore, errors in
for example). This relationship is established by predicting individual sensors are statistically dependent on each other,
measurements on the basis of a model and calculating a and this must be taken into consideration as well (see e.g.
predicted range within which the measurements should fall [11]).
(“gating”). Afterwards, the measurements are assigned to the Table I summarizes the most important aspects of this
track which matches best, whereby different optimization discussion. The results suggest that a hybrid architecture
strategies may be used (see e.g. [3], [13]). incorporating the advantages of both approaches is desirable.
In addition to information about the (dynamic) states of an B. Feedback from the Applications to the Sensors
object used in an application, additional information If the characteristics of sensors are variable (e.g. field of
accumulated by the sensor may be used during the association view, detection rage), it is important to ensure defined
process (e.g. bitmap information about an object in a video feedback from applications to sensors. Ultimately, one can
sensor). This will result in fewer errors in the association only proceed from the specific application in order to
process as compared to relying on state estimations only, as determine which aspect of the environment is of interest at
more information is used in this step. any given time and thus to define the sensor configuration.
TABLE I ApplicationLevel
Application Level
Application Level
CENTRALIZED VERSUS DISTRIBUTED DATA PROCESSING Application
Requirement/
Requirement centralized decentralized L2 View Feedback

No loss of information Fusion Level


+ - Coordination of
Requirements
High degree of accuracy + - Fa F1 ... Fn

Consistent model assumptions + - L1 Object-Class


Requirement
Low bandwidth - + Association II, Classification, for Sensor
Verification
Homogenous processor load - +
L1 View
Easily modifiable - +
Sensor Level
Translation/Compression Translation

Sensor-
However, different applications can pose conflicting Association I configuration

requirements, as in the case of a variable detection range of a Sensor – Signal Processing

medium- to close-range sensor used in such applications as Fig. 5. Modular System Architecture for Fusing Environment Sensor Data
Stop&Go, Parking Assistant or airbag preconditioning.
object (such as velocity, position, etc.) alone, but also on such
These conflicts have to be resolved. In this process, the
sensor-specific information as the reflectivity measured by a
applications must be informed, whether a requirement has
laser scanner or image-data detected by a video sensor. The
been met or not, in order to be able to react accordingly.
association step may be skipped for sensors in cases in which
C. Classification of Objects an association step at the sensor level is undesirable (due to
In addition to the straight estimation of state variables for a the sensor principle or to cost considerations). Data that
recorded object (e.g. velocity, etc.), classification of the cannot be associated (e.g. objects which are detected for the
objects (i.e. identification of the type of an object.) is first time) will be sent to the fusion level without association.
important for both applications and state estimation The next step at the sensor level is the
algorithms. translation/compression, through which data is converted into
In addition to enabling applications to derive longer-term a generalized view (so called “L1 view”). Wherever possible,
strategies, this information also makes it possible to select data in L1 view has not yet been subjected to model-based
appropriate model assumptions for use in estimating state filtering. This takes place at the fusion level only, thus
variables (a traffic sign has different physical constraints in guaranteeing that the model remains uniform over the entire
comparison to a car, for example). fusion process.
Moreover, a long-term prediction of the behavior of an Translation makes it possible to standardize data from
object that is based on more than physical constraints alone is sensors based on different principles or from different
possible only if the type of object is known (the behavior of generations of sensors. However, as it is not possible to obtain
an observed pedestrian may differ from that of a detected a standardized view of all sensor data conceivable (new
cyclist). principles might still be developed), the L1 view must be
capable of expansion.
VI. MODULAR SYSTEM ARCHITECTURE Separation between pre processing and filtering will
The following section describes a system architecture actually not be possible in every case. Thus a model is
developed on the basis of the results discussed above (see also required even for the interpretation of radar data readings, for
[9]). It is shown in Fig. 5. example. If separation is not possible, the minimum
requirement should be that potential falsification and
misinterpretation of the data based on model assumptions is
A. Sensor Level minimized and filtering is performed only to the extent
The sensors represent the lowest level. Raw data signal necessary for a particular identification.
processing may be preconditioned with the aid of data traced B. Fusion Level
back from the fusion level. This is useful, for example, in
video image processing (definition of search areas) as a means Data entering the fusion level is first verified to minimize
of reducing computational load (see for example [12]). errors. Classification is performed using all available sensor
The measured data is then passed on to the association information and data already obtained from a track. Data
function. Here, whenever possible, the measured data is which is not associated to a track so far is associated at this
associated to the previously established tracks, which are stage, based on state variables only. Data that cannot be
traced back from the fusion level. With the association step associated and which has passed the verification step is used
established at the sensor level, additional information from the to establish a new track.
sensor that is not reported to the higher-level data processing There are two major advantages to incorporating a second
function at the fusion level can be used for this task. In this association step at the fusion level, in addition to first step at
way, association is not based on the estimated states of an the sensor level. First of all, the association step may be left
out at the sensor level in order to reduce costs; secondly, there
is a chance that two sensors may detect a new object at the of existing applications and so allows keeping already tested
same time and report it to the fusion level as a new track modules in the ongoing development process.
(depending on the timing of the sensors). If an association step The concept is open with regard to both sensors and
is incorporated at the fusion level resulting multiple track applications. The communication bandwidth between the
initializations will be minimized. different layers is reduced significantly as compared to that
The associated and classified measurements are then passed required for raw data transmission, while error propagation is
on to a filter bank, which consists of different model-based minimized and remains transparent.
filters customized for different types of applications and one The architecture itself is independent on the hardware
filter especially designed for the requirements of the topology. It might be implemented either on a single processor
association. This filter works separately from the filters system in which the different layers correspond to different
adapted for the applications needs by using purely application software modules or on a multiprocessor system in which each
independent object properties. instance of the sensor level or the application level
Each application dependent filter can be (statically) corresponds to one piece of hardware and the L1 and L2
assigned either to one individual application or to a group of views are transmitted via a bus system.
applications. In this way, multiple implementations are Defining a standardized L1 and L2 view acceptable to
avoided and sufficient flexibility remains for the various multiple manufactures would then make it possible to use
requirements of different applications. hardware (and software) modules from different suppliers
Filter parameters can be influenced on the basis of interchangeably.
identified object type and dynamic application requirements In the next step, the architecture will be implemented in a
(which must be coordinated in the case of a multiple test vehicle during the project PRORETA [14]. The two main
allocation). The result of the estimation process is the so hypotheses (“Different application groups need different
called “L2 view”, which is passed on to the application level model assumptions in order to work optimally” and “Data
and also traced back to fusion and sensor levels. association at the sensor level is more reliable than data
In contrast to the L1 view, the L2 view is independent of association at the fusion level”) of the system architecture will
sensor principles. Further development of this view will be analyzed in a real environment.
depend solely on new application requirements.
REFERENCES
C. Application Level
[1] Becker, J.-C., “Fusion der Daten der objekterkennenden Sensoren eines
The applications receive data in the L2 view and can select autonomen Straßenfahrzeugs.” Dissertation, TU-Braunschweig,
the filter properties (models), which best match their Fortschr.-Ber. VDI Reihe 8, Nr. 948, VDI Verlag, 2002.
requirements. From the standpoint of a given application, [2] Klein, L. A., “Sensor and Data Fusion Concepts and Applications,” 2nd
edition, SPIE, 1999.
there is only one “virtual sensor”, which delivers data in [3] Bar-Shalom, Y., Xiao-Rong, L. “Multitarget-Multisensor Tracking:
accordance with the requirements of the specific application. Principles and Techniques,” 3rd printing, YBS, 1995.
The process through which information is generated remains [4] Sparbert, J., Dietmayer, K., Streller, D., “Lane Detection and Street Type
Classification using Laser Range Images,” in Proc. IEEE Intelligent
transparent. Transportation Systems Conference, 2001.
The properties of the virtual sensor (e.g. the dynamics of [5] Vukotich, A., Kirchner, A., “Sensorfusion für Fahrerassistenzsysteme,”
the data filtering process, the field of view of the “virtual in VDI-Berichte 1646, pp. 857-875, VDI Verlag, 2001.
[6] Lou, R.C., Kay, M.K., “Multisensor Integration and Fusion in Intelligent
sensor”) can be influenced dynamically by the applications. Systems”, in: Autonomous Mobile Robots Volume 1, IEEE Computer
The requirements are prioritized at the fusion level and Society Press, Los Alamitos, California, 1991.
converted into generalized sensor requirements, which are [7] Joerg, K.-W., “Echtzeitfähige Multisensorintegration für autonome
Mobile Roboter,” Mannheim, etc., BI-Wiss.-Verl., 1994.
again formulated in a standardized language (as in the L1
[8] Darms, M., Winner, H., „Fusion von Umfelddaten für
view) and converted into a special sensor configuration at the Fahrerassistenzsysteme“, in Stiller, C., Maurer, M., Workshop
sensor level. Fahrerassistenzsysteme FAS2003, pages 13-16, 2003.
Because different requirements may conflict and thus [9] Darms, M., Winner, H., “A General System Architecture for Fusion of
Data from Environment Sensors for ADAS” ITS 2005, Hannover,
remain unfulfilled for some applications, feedback from the 03.06.2005.
fusion level to the application level enables the application to [10] Randler, M, Wilhelm, U., Lucas, B., “Anforderungen von Komfort- und
react accordingly. Sicherheitsfunktionen an die Umwelthypothese der Umfeldsensorik,” in
Stiller, C., Maurer, M., Workshop Fahrerassistenzsysteme FAS2003, pp.
42-44, 2003.
VII. CONCLUSIONS & SUBSEQUENT STEPS [11] Bar-Shalom, Y., Xiao-Rong, L. “Estimation with applications to tracking
and navigation,” Wiley, 2001.
The definition of a system architecture for sensor data [12] Franke A.G., Levi, P. “Robust Vehicle Tracking Fusing Radar and
fusion of environment sensor data for Advanced Driver Vision,” International Conference on Multisensor Fusion and Integration
Assistance Systems must account for the fact that processing for Intelligent Systems, 2001.
[13] Hall, D.L., McMullen, S. “Mathematical Techniques in Multisensor Data
of sensor data is application dependent.
Fusion,” 2nd edition, Artech House, 2004.
The system architecture presented in this article takes this [14] Project Homepage PRORETA, www.iat.tu-darmstadt.de/proreta,
into account. Furthermore it allows the addition of new 22.04.2005.
applications without any modification of the underlying filters

Você também pode gostar