ACCIDENT ANALYSIS BASED SAFETY FUNCTION DEVELOPMENT

Transcription

ACCIDENT ANALYSIS BASED SAFETY FUNCTION
DEVELOPMENT
Author 1
Roland
Galbas
Author 2
Dr. Thomas
Heger
Author 3
Lisa
Blank
Author 4
Dr. Michael
Fausten
Author 5
Andreas
Georgi
Robert Bosch GmbH
Robert Bosch GmbH
Robert Bosch GmbH
Robert Bosch GmbH
Robert Bosch GmbH
Combined Active and
Automtoive Electronics
Corporate Research
Combined Active and
Corporate Research
Passive Safety
Driver Assistance AE-
CR/AEV1
Passive Safety
CR/AEV1
CC/PJ- CAPS
DA/ESA3
Robert-Bosch-Strasse 2,
CC/PJ- CAPS
Robert-Bosch-Strasse 2,
Robert-Bosch- Allee 1,
Daimlerstraße 6, 71229
71701 Schwieberdingen
Robert-Bosch- Allee 1,
71701 Schwieberdingen
74232 Abstatt Germany
Leonberg Germany
Germany
74232 Abstatt Germany
Germany
Roland.Galbas@de.bosch.
Thomas.Heger@de.bosch.
Lisa.Blank@de.bosch.com
Michael.Fausten@de.bosc
Andreas.Georgi@de.bos
com
com
h.com
ch.com
ABSTRACT
The development of modern automotive safety functions is strongly driven by reducing
development effort, component costs, tight scheduling and especially the expectation of high
function efficiency.
As improving safety is a shared vision of governmental authorities and industry, the industry
is called upon to exhibit proven evidences of function efficiency – even before a dedicated
function works in field.
The given challenges can be mastered by involving the latest results of accident analysis into
the development process. The extensive use of in-depth accident data enables optimized
function design with focus on efficiency and related component costs. Hence, an accident
analysis based methodology for the development of active safety functions is derived. The
power of this method will be demonstrated by the application on an exemplary use case.
MOTIVATION
Usually the development process is characterised by recursive steps within the prototype
phase. This is typically caused by the complex interdependencies of product liability,
heterogeneous use cases and related component costs. A more efficient development process
can be achieved by answering the following questions:
-
How to focus precisely on accident situation with statistical relevance?
-
How to shift recursions from prototyping to simulation phase (Frontloading)?
-
How to derive requirements for components in order to balance costs and performance
optimally, all this at an early stage of development?
Accident analysis is the key to a development method for safety functions which copes with
the given challenges
1
MAIN PROCESS OF FUNCTION DEVELOPMENT
An appropriate method to address the above mentioned challenges is the accident data based
development process. The process mainly consists of two phases:
Phase 1: Macro-data is analyzed in order to identify relevant accident scenarios. Accident
kinds, causes and the function placement into the accident chronology (e.g. according to the
ACEA model) are used to support the function design.
Phase 2: The function draft is implemented into a simulation environment. The reproduction
of real world accident scenarios thus enables recursive development. Within this phase
detailed in-depth accident data is used to support the well known development methods.
PHASE 1
A) MACRO DATA BASED ASSIGNMENT OF SAFETY FUNTIONS
The distribution of accident kinds and related causes allows focusing on relevant function
approaches. Figure 1 shows the statistical relevance of accident kinds and accident causes in
Germany. The top event for fatalities can be widely adressed by instsllation of an ESP®. Accidents
with oncomming traffic might be adressed by means of C2x related functions. Since the
accident situation on crossings is rather complex, in this paper we choose this accident kind
for the introduction of the development method.
Leaving the carriageway to the right or left
Accident
kinds
Source: GIDAS, 2004, BaSt 2004
10%
Fatalities distribution of accident kinds
Collision with another oncoming vehicle
6%
Collision with another vehicle which
turns into or crosses a road
Collision between vehicle and pedestrian
Collision with another vehicle which is
moving ahead or waiting
Others
36%
12%
14%
22%
5% 3%
2%
8%
23%
14%
47%
Accident
causes
1%
36%
13%
48%
30%
35%
35%
priority, precedence
turning, U-turn, reversing, entering
the flow of traffic, starting of the
edge of the road
speed
Fatalities distribution of causes for collision with
another vehicle which turns into or crosses a road
speed
improper driving
overtaking
driving fitness
technical or maintenance faults
others
Fatalities distribution of causes for collision with
another oncoming vehicle
speed
other mistakes by the driver
driving fitness
improper driving
others
Fatalities distribution of causes for collision while
leaving the carriageway to the right or left
Figure 1. Fatalities vs. accident kinds and causes (Germany)
B) MACRO DATA BASED FUNCTION APPROACH
The function design highly depends on the operating phase within the ACEA accident
chronology. Functions acting in early stages of the accident chronology usually address rather
the accident causes than the kinds (e.g. speed warning). Functions acting in late phases of the
crash chronology typically focus more on accident kinds (e.g. side airbag deployment).
2
Influence of causes
Priority,
precedence 7%
Turning, U-turn,
reversing,… 4%
Collision with
another vehicle
which turns into
or crosses a
road, (14%)
Speed 3%
Accident kind
Accident
causes
Accident kinds, causes of fatalities Germany, Source: GIDAS, 2004, BaSt 2004
Others < 1%
Risk
avoidance
Increased
risk
High
risk
Crash
inevitable
Active safety
Incrash
After
crash
Passive safety
Figure 2. Fatalities in Collision with another vehicle which turns into or crosses a road
Figure 2 shows the example of a “collision with another vehicle which turns into or crosses a
road” and its related causes, equalling 14% of fatalities in Germany. Hence, the influence of
the accident causes along the accident chronology has to be analyzed and carefully checked
against the initial function approach.
PHASE 2
SIMULATION BASED SYSTEM DEVELOPMENT WITH IN-DEPTH ACCIDENT
DATA
After selecting in phase 1 the function focus and its alignment to the accident timeline, the
function and the required components have to be designed and tailored to the real world use
case.
In preparation of phase 2 the accident scenarios are chosen according to the envisaged use
cases by several criteria like accident kinds, accident types, collision types or others. After the
selection process the chosen accident scenarios have to be statistically aligned to official
national accident databases (in Germany provided by the “Federal Highway Research
Institute of Germany”- BAST). This alignment is done under consideration of the selected
criteria.
From the selected in-depth accident data, the use cases are transferred to a simulation tool,
were function algorithm and system component specification can be optimized for maximum
real world performance and best cost-benefit ratio.
The availability of in-depth accident data is essential for this process. For Germany data from
the German In-Depth Accident Study (GIDAS) with typically up to 3500 items for each
accident are available. Further regional in-depth databases can be used in order to achieve a
more global function shape. For data transfer from the accident data bases to the simulation
tool, an automatic tool should be used for workload reasons.
Only now, in a third step, recursive optimizations adapt the selected system to variable
boundary conditions, using first samples.
3
As a result, precisely defined requirements components are obtained, balancing costs and
performance. In parallel a proven efficiency of the function is provided. Thus many recursions
could be shifted from prototyping to simulation phase (Figure 3).
step 1 - accident transfer
step 2 - function integration
¾
accident
database
¾
Simulation
environment
¾
Function
approach
step 3 - Recursive Approach
¾ Simulate & Analyze main objectives
¾ Adjust function, components, parameter
Select hardware and first test
Figure 3. Base and process of function development
Below, the development process shall be exemplified by designing a system for avoidance or
mitigation of a “collision with another vehicle which turns into or crosses a road”.
DEVELOPMENT STRUCTURE FOR SIMULATION BASED
FUNCTIONS DEVELOPMENT
For the system development according to step 2, two main development objectives can be
identified: Objective 1: Function algorithm analysis in accordance to the selected use cases
and related features. Objective 2: Sensor parameter analysis in accordance to the selected
use cases (Figure 4). The discrimination of objectives together with a careful selection of the
accident items transferred to simulation enable a considerable reduction of effort in terms of
accident data transfer.
Sensor parameter analysis
- Reduced level of detail
- Huge number
Function algorithm analysis
- High level of detail
- Moderate number
Transfer of acc. To full sensor simulation
Theoretical calculation: Sensor and Situation
Validation
Huge num. of relevant accid.: street angle, veloc.
Object – Identification & classification
Sensor parameter definition: angle, distance,..
Ego Motion of Veh. by e.g.ESP
Transfer of sensor parameter
Moderate number of accidents
Object- Identification, -classification
-> Probability of Collision
Figure 4. Development structure
4
Function algorithm analysis
The development of the function algorithm demands highly detailed real world scenarios in
order to reveal all sensor relevant factors and their influences on the concerned system.
Exactly reconstructed road- and accident conditions and their comprehensive validation is the
base for reliable statements of this analysis (Figure 5). It has to be ensured that every targeted
use case is covered.
Figure 5. Integration and validation of accidents
Sensor parameter analysis
The sensor parameter analysis targets amongst others two main
characteristics of surround sensing systems: field of view and
detection range. Weighty statements are only achieved if all
influencing items - in the given example velocity of vehicles, view
angles including occluding objects, friction of road, road dimension
and some others - are exactly reconstructed (Figure 6).
Fzg.2
d2
dα1
α
d1 βt1
γ
In doing so, the number of chosen accident scenarios needs to be
representative and aligned to the respective macro data base.
As a result a data base providing the required sensor parameters for
enabling accident avoidance or mitigation within the targeted use
cases is obtained.
v2
v1
Fzg.1
dobj
T=t1>t0
Figure 6. Reduced item sensor analysis
SENSOR SELECTION AND DESIGN
Based on the results of the use case analysis and the required sensor parameters (FOV) an
adequate sensor concept for the system has to be designed and assessed. Since in the
automotive technology field various sensor principles and sensors exist - generally speakingthis process has to be carried out for each of them.
As an example we show some results of the assessment of a video sensor.
5
In consideration of the relation between field of view and detection range a sensor specific
coverage graph as shown in
Figure 7 can be derived from the accident data base.
Independent variables for this characteristic curve are:
¾ number of imager columns of the camera
¾ focal length of the camera
¾ expected size of most relevant objects in the world
¾ minimum required object size for detection (in pixels)
¾ system setup time (system response plus driver reaction time)
¾ minimal functional impact for accident mitigation or avoidance
percentage of avoidable accidents for different fields of view
120,00
100,00
[%]
80,00
+/- 55°
60,00
40,00
+/- 23°
20,00
0,00
FOV (field of view ) [+/- °]
Figure 7. Coverage graph
The derived coverage graph in
Figure 7 indicates, for which FOV which percentage of
accidents can be addressed. For example a FOV of about ± 55° results in coverage of 80% of
the analyzed accidents. A FOV of a typical front looking LDW-camera (about ± 23°) yields to
a coverage of about 35% of avoidable accidents. Thus, the expected system performance for
any video sensor setup can be specified.
SYSTEM EVALUATION
When looking at the scenario of the evaluated accidents, it becomes clear that for a video
sensor an algorithm for object detection is needed, which is independent of shape and position
of the critical objects within the image. In addition to that it would be very helpful if the
algorithm was very sensitive to motion within the image, because the possible collision object
in the exemplary accident kind usually moves laterally towards the driving corridor of the ego
vehicle. Hence, the Bosch Dense-Flow algorithm can be applied. Measurement results can be
seen in Figure 8. (2),(3).
6
Figure 8. Dense optical flow field (Bosch Dense-Flow)
One inherent problem when dealing with accidents is that it is very difficult to set up a
realistic test scenario. To test the performance of the detection algorithm, a selection of
accidents from the GIDAS data base was simulated in high detail. The set-up of an accident
scenario in the simulation environment was done by combining accident data with accident
images resulting in generated synthetic videos of the accidents including all the relevant
objects in accordance to the real world data. An example of such a simulated scene is shown
in Figure 9.
Figure 10 shows the detection performance of the algorithm on that simulated scene. This
approach allows testing the algorithm under realistic conditions such as distance, speed and
visibility without any risk of physical or even human damage. In addition to that, by means of
this procedure any accident can be simulated repetitively in exactly the same way in order to
optimize the detection algorithm.
Figure 9. Sequence of a reconstructed and simulated real accident
Figure 10. Detection of crossing vehicle in simulated scene
REAL WORLD PERFORMANCE
After finishing the system design and its assessment the simulation results need to be verified
against real world performance of the system. The evaluation of the real world performance
7
requires real world measurements for the considered sensor concept with focus on false
positive and true negative rate.
For the verification of a video sensor, a standard video camera for vision based driver
assistance applications was used in real world scenes.
Figure 11 shows such a scene with the detection of a crossing vehicle.
Figure 11. Real world performance of the detection algorithm (parameter setting tuned for
demonstration purposes)
The Time To Collision (TTC) calculation which enables a warning strategy base on two main
data sources
1) The ego motion estimation - derievd from video data and ESC data
2) The object relation estimation based on various approaches
- Trigonometrical distance calclations presuming a flat road and using the camera
angles, and the target vehicle bottom line
- Object scale change over time
- Object direction estimation based on the optical flow measured on the surface of the
targeting objects (Figure 12)
-
Figure 12 optical flow measured on the surface of identified object (simulated artificial scene)
8
SUMMARY AND OUTLOOK
A development method was presented that allows an optimal design of a vehicle safety
system based on real word accident data. Real world performance is optimally balanced to
system cost in the early definition phase. Recursions during the specification process and thus
development costs are minimized. The expected system performance is staging for marketing
and management decision processes (Figure 13).
Moreover, the method provides an excellent knowledge base when tailoring a safety system to
OEM specific needs
Method – supporting safety function development
Î Supports decision for system
Î Supports system design
Î Time and costs in development saved
Î Function effectiveness proven
Î System dependencies assessed
Figure 13 support method safety function development
ACKOWLEDGEMENT
Thanks to the GIDAS Cooperation for sharing and supporting our vision of efficient accident
reduction.
REFERENCES
(1) GIDAS (German in-depth Accident Study)
(2) F. Stein, Efficient computation of optical Flow, In Proceedings of the 26th DAGM Pattern
Recognition Symposium. Tübingen, Germany: Springer, August 2004.
(3) R. Hartley and A. Zisserman, Multiple View Geometry in computer vision, 2nd ed.
Cambridge Press, 2003.
9
Unfallanalyse & Sicherheitsfunktionen
Methodische Nutzung der Unfallanalyse
zur Entwicklung von Sicherheitsfunktionen
Roland Galbas
Robert Bosch GmbH
1
CC/EVM-Galbas | 03/03/2008 | © Robert Bosch GmbH 2007. All rights reserved, also regarding any disposal, exploitation, reproduction, editing,
distribution, as well as in the event of applications for industrial property rights.
Eckpunkte der Systementwicklung
Funktionszielstellung
¾ Statistische Nutzenrelevanz
Î System Design
¾ Anforderungen an Komponenten basierend auf
Nivellierung von Kosten und Performance
Î Effizienz im Entwicklungsprozess
¾ Verschiebung der Rekursionen aus
Prototypenphase zur Simulationsphase
(Frontloading)
Î
Î
2
Instrument - Intensive Nutzung der Unfallforschung
Prozess
Phase 1
Makrodatenanalyse
Î Wirkfeld und Fokus – Warum und Was?
Î Unfallchronologie – in welcher Phase?
Phase 2
Detaildatenanalyse
Î Reproduktion realer Unfallszenarien
Î Implementierung der Rohfunktion in
Simulationsumgebung
¾ Rekursive Entwicklung des Funktion
3
Phase1
Makrodatenanalyse
Abkommen von der Fahrbahn
Kollision mit entgegenkom. Fzg.
Kollision mit anderem Fzg. das einbiegt
oder kreuzt
Unfallarten
Quelle: GIDAS, 2002- 2004,
10%
Todesfälle Verteilung von Unfallarten
6%
12%
36%
Kollision zwischen Fzg und Fußgänger
Kollision mit anderem Fzg, das vorausfährt
oder wartet
14%
22%
Andere
5% 3%
Unfallursachen
8%
Î
4
2%
1%
14%
23%
47%
36%
13%
30%
48%
35%
35%
Vorfahrt
Abbiegen, Wenden, Rückw.
Fahren, Ein- + Ausfahren
Geschwindigkeit
Geschwindigkeit
Geschwindigkeit
Straßenbenutzung
Überholen
Verkehrstüchtigkeit
Techn.. Mängel, Wartungsmängel
Andere
Andere Fehler beim Fzg.führer
Verkehrstüchtigkeit
Straßenbenutzung
Andere
Nutzen – Fokus auf statistisch relevanten Funktionen
Phase1
Î
Makrodatenanalyse
Beispiel: Betrachtung der Unfallchronologie
Unfallursachen
Einfluss der Ursachen
Vorfahrt 7%
Turning, U-turn,
reversing,… 4%
Speed 3%
Others < 1%
Risiko
Vermeidung
Erhöhtes
Risiko
Hohes
Risiko
Aktive Sicherheit
Î
5
Unfall unvermeidbar
Im Unfall
Nach Unfall
Passive Sicherheit
Nutzen – Schärfung Designfokus
Kollision mit
anderem Fzg.
das einbiegt
oder kreuzt
(14%)
Unfallarten
Unfallarten, Unfallursachen für Todesfälle in Deutschland, Quelle : GIDAS, 2002- 2004,
Phase 2 - Detaildatenanalyse
Sensor /Aktor
Fzg.1
System Design
d2
dα1
α
d1 βt1
γ
v1
dobj
Î
Fzg.2
v2
T=t1>t0
Î
Blickwinkel (FOV)
Entfernung
+
Algorithmus
Datenbasis Funktionsentwurf
Rekursiver Ansatz
Algorithmus + Szenen
Î Sensorparameter
Î
Î
6
Î
Unfallsimulation
¾ Simulation & Analyse der
Kernfragestellungen
¾ Anpassung Funktion, Parameter,…
Key benefit: balancing costs and performance
Sensor Auswahl and Design – Beispiel Video
Sensor Parameter
Î FOV, Abstand,…
Funktionaler Rahmen
Î Min Objektgröße,
Î Reaktionszeit,…
Effektivität des Systems
Î z.B. Abdeckung vermeidbarer Unfälle
Prozentuale Abdeckung vermeidbarer Unfälle für
verschiedene Blickwinkel (FOV)
100,00
80,00
+/- 55°
[%]
Fzg.2
d2
dα
1
α
v2
60,00
40,00
20,00
+/- 23°
0,00
d1 βt1
γ
Blickwinkel - FOV [+/-°]
v1
Fzg.1
dobj
T=t1>t0
Î
7
Nutzen – Direkter Link: System
Performance
Real-world & Systemevaluierung
Sensor
Î Video (FOV)
Fzg.2
d2
dα1
α
v2
Algorithmus
Î Power
Flow
Unfalldatensimulation
Î Use Case Performance
d1 βt1
γ
v1
Fzg.1
dobj
T=t1>t0
Prototyp and Test
Î Real-world Performance
Î Fehlwarnungen
¾ Weiterentwicklung
Sequence
8
Systemevaluierung
Zusammenfassung
Methodische Nutzung der Unfallanalyse zur
Entwicklung von Sicherheitsfunktionen
Unterstützung in Systementscheidung
Î Unterstützung im Systemdesign
Î
Kosten- und Zeitersparnis in Entwicklung
Î Nachweis der Effektivität der Funktion
Î Systemabhängigkeiten sind bewertbar
Î
9
Danke für Ihre Aufmerksamkeit!
10

ACCIDENT ANALYSIS BASED SAFETY FUNCTION DEVELOPMENT

Transcription

Similar documents

Berlin – Stadt des Designs

zulassungsschein - TES