Skip to main content
Erschienen in:

Open Access 2022 | OriginalPaper | Buchkapitel

2. Findings from Traffic Accident Analysis

Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.

loading …


With regard to safe product development in the dilemma between Artificial Intelligence, ethics and legal risk, Thomas Winkle provides a meta-analysis for safety assessment using accident data to demonstrate potential safety benefits and risks. Thomas Winkle also refers to the disasters of the Chernobyl and Fukushima nuclear power plant accidents. They mark changes in acceptance about the use of nuclear energy. Comparisons are made between the global mortality rates of females and males on traffic mortality in relation to the life expectancy of various countries around the world, from Sierra Leone with the lowest life expectancy to Japan with the highest. The probability of dying from a traffic accident is highest in Venezuela and Sierra Leone. Another overview addresses the global mortality rate with exemplary causes of death in contrast to the ISO 26262 requirements of the Automotive Safety Integrity Level "ASIL D" and a hardware failure rate of less than 1 * 10-8 1/h. Furthermore, he uses traffic accident data examples of safety-enhancing automated vehicle systems with a low degree of automation that are already available on the market. For testing methods to develop and validate safe automated vehicles with reasonable expenditure, the author recommends combining worldwide traffic accident-, weather-, vehicle operation data and traffic simulations. Based on these findings, a realistic evaluation of internationally prospective, and statistically relevant real-world traffic scenarios as well as error processes and stochastic models can be analyzed (in combination with virtual tests in laboratories and driving simulators) to control critical driving situations.
This chapter starts with findings and limits of accident investigation regarding potential safety-enhancing vehicle systems with low degrees of automation.
Contents of this chapter were already prepublished within the springer book: Autonomous driving – technical, legal and social aspects (Winkle, Safety Benefits of Automated Vehicles: Extended Findings from Accident Research for Development, Validation and Testing, 2016a).
So far, no sufficient experience with series applications of fully automated vehicles has existed. A safety prognosis of such features depends on assumptions regarding market penetration and technological progress.
Therefore, based on his work experience, the author recommends combining area-wide traffic accident-, weather-, and vehicle operation data as well as traffic simulations in order to develop, test and validate safe automated vehicles with reasonable expenditure.
The aim is to focus on the essentials and to validate using a scenario catalogue. Few tests under special conditions replace many simple tests. Taking into consideration human and machine perception, these findings result in a realistic evaluation of internationally and statistically relevant real-world traffic scenarios as well as error processes and stochastic models. These, in combination with virtual tests in laboratories and driving simulators, can be analyzed to prevent critical driving situations.

2.1 Motivation

Since the beginning of the millennium, automobile manufacturers have made active steering-assistance systems (Lane Keeping Assistance Systems – LKAS) in combination with active distance keeping (Adaptive Cruise Control – ACC) for series production vehicles available. The combined functionality was introduced into the Japanese market for right-hand drive vehicles such as the Nissan Cima (2001) and the Honda Inspire (2003). Since then, partially automated driving (see Ch. 2.2) of up to 20 seconds has been possible under the driver’s supervision when using both assistance systems (author’s test drives in 2003). German manufacturers, starting with the VW Passat CC (2008), have been selling active steering systems in selected models as an optional feature (Katzourakis, Olsson, Lazic & Lidberg, 2013). In December 2021, the German Federal Motor Transport Authority granted the world’s first type approval for an Automated Lane Keeping System (ALKS) from Mercedes-Benz. The basis for the Drive Pilot in the Mercedes S-Class and EQS (2022) is UN Regulation Number 157 (Kraftfahrtbundesamt KBA, 2021).
In times of increasing market penetration of active safety systems statistics by the Federal Statistical Office of Germany have shown a decrease of road accident fatalities: While 21.332 people died in road accidents in Germany in the year 1970, the number was reduced by more than six times to 2020 with 2,719 fatalities (Statistisches Bundesamt 2018). This is even more significant as at the same time driven mileage increased by almost 30 percent (251 billion kilometers in 1970, 736 billion kilometers in 2018 (Kraftfahrtbundesamt). Among the remaining accidents are some that might have been prevented by automated vehicle functions. Potential safety benefits can be determined on the basis of accident data, namely the fall of accident-related fatalities. Examples given in this book demonstrate the possibilities and limits of analyzing this data.
Various organizations carry out traffic accident research all over the world. This encompasses the subfields of accident surveys/statistics, accident reconstruction, and accident analysis (Kramer, 2013). The basis for accident research in Germany is investigation, carried out by the police. Additionally, other institutions carry out their own accident research, such as the Traffic Accident Research Institute of TU Dresden GmbH (Verkehrsunfallforschung, or VUFO) and the Hannover Medical School, as well as vehicle manufacturers and the German insurance industry. A comprehensive source of data is the investigation of accidents at the scene, which are also statistically recorded and evaluated according to certain weighted characteristics. Acquired data can be used for the safety-enhancing further development of vehicle automation. The following chapters exemplarily demonstrate automated vehicles’ potential safety benefits, limits of findings and predictions resulting from accident data collections.
The following chapters focus on two questions, using specific examples from accident research:
  • How significant are analyses and findings from road accident research for the introduction of connected automated vehicles?
  • How can potential safety benefits of automated vehicles be proven?

2.2 Categorizing the Levels of Driving Automation

To illustrate the potentials and limits of accident data analyses, three categories for levels of driving automation (concerning the degree of vehicle guidance) will be used. This categorization is derived from a BASt-project publication “Legal consequences of an increase in vehicle automation” (Gasser et. al. 2012), which lists two further categories. Their five degrees of automation start with conventional vehicle guidance, called “driver only”, where the driver is constantly responsible for the vehicle’s longitudinal and lateral motion. The classification continues with driver assistance (“assisted”) and partial automation (“partial automated”), with permanent driver supervision. Lastly, the levels of highly automation (“highly automated”) and full automation (“fully automated”) permit humans to stay out of the vehicle guidance process some or all the time (Gasser et. al. 2012). Vehicles currently on the market are neither highly nor fully automated. As a consequence, no accident data exist regarding these categories, which therefore will play no role in the examples below.
In order to give a complete overview another two classifications are mentioned: Similar to the BAST project, five levels were defined by the American NHTSA agency (National Highway Traffic Safety Administration, 2013). Subsequently, the SAE International (formerly Society of Automotive Engineers) developed six distinctions in its SAE J 3016 standard and describes their minimum requirements. In ISO/SAE PAS 22736, published since 2021, these six levels have been adopted. They have been valid since January 2014 and commonly used today. These levels correspond to the BASt levels published previously in 2012, with two differences. Not only the names of the levels are different but SAE adds level 5 (full automation): at this level the automated driving system performs the complete driving task under all conditions a human driver can manage (Society of Automotive Engineers, 2014); (see Fig. 2.1). The technical definition “fully automation” is also described under the term autonomous driving technology and includes a variety of possible applications and characteristics (e.g. Interstate Pilot, Valet Parking, Vehicle on Demand, Driver for Extended Availability) (Wachenfeld et. al., 2016; Donges, 2016). A total of three instance groups (“internal” e.g. adult or underage passengers, disabled persons, “the driving robot” and “external” e.g. authorities, police) can take over the driving of the vehicle.
Fundamental questions to the developers are:
  • At what level of vehicle guidance does an internal, external group or the autonomous vehicle itself have the ability to intervene?
  • At what level of vehicle management does an internal, external group or the autonomous vehicle itself have the authority to intervene?
  • Which instance is dominant in the conflict of simultaneous intervention?
  • How is the hierarchy between the instances defined?
  • Is the autonomous vehicle allowed or does it have the possibility to disregard applicable rules in order to avoid greater damage?

2.3 Accident Data to Demonstrate Potential Safety Benefits and Risks

Basically, automotive technology has always been considered as a technology with undesired side effects. An unambiguous understanding of acceptable risks that can be taken as a basis for decisions on automated system designs is a prerequisite for a safe development process.
Where do relevant risks caused by automated driving come from?
First of all, safety-related failures caused by hardware (random failures and design errors) are possible. Furthermore, software errors (design errors and inadequate quality assurance) will continue to gain significance for increasing importance. Such issues have been discussed for many years within automotive manufacturers and suppliers. Many new standards have been established to ensure traffic safety over the last years.
Behind all these activities however, a basic question always has to be answered: What is an acceptable risk of automated driving technologies that can be determined and evaluated? People take risks when they have personal control. Is the assessment of risk based on frequencies or probabilities? How is the risk perceived? Will it be accepted or not?
In general, there is a strong tendency to assess risks based on individual cases. A single accident can be an opinion-forming event. On April 26, 1986, a unit of the Chernobyl nuclear power plant in Ukraine exploded. About 25 years later, the reactor cores of three reactors at the Fukushima Daiichi nuclear power plant in Japan melted on March 11, 2011. Although the two disasters are not comparable, both Chernobyl and Fukushima have released massive amounts of radioactive material. Two clearly different reactor types were affected. Block 4 at Chernobyl was a water-cooled and graphite-moderated reactor. A combination that can trigger uncontrolled chain reactions, which occurred in the case of Chernobyl. The accident was caused by an experiment carried out by the operating crew, which got completely out of control. The plan was to simulate a complete power failure in order to show that the turbine would still supply sufficient power even after the reactor had been shut down, so that the time required for the emergency units to start could be bridged.
In Fukushima, the reactors from the Tokyo Electric Power Company (TEPCo) stand on granite foundations. They are surrounded by steel and concrete structures. Trigger of the accident in Japan was a huge earthquake. As a consequence, the subsequent tsunami flooded the coastal nuclear power plant, which caused the power in the high-voltage grids to fail. Therefore, the systems ran on emergency power until the tsunami shut down the emergency diesel engines. Batteries remained, but were exhausted after a few hours. From then on, no more cooling water of the Reactor Coolant System (RCS) was pumped over, so that the reactor cores and the fuel elements stored in the decaying ponds of the piles overheated.
So far, the two accidents have been the only ones to which the highest level on the international INES reporting scale has been assigned. The INES (International Nuclear Event Scale) is used to assess accidents in nuclear facilities.
The Chernobyl and Fukushima disasters mark changes in acceptance with significant turning points in environmental policy and in the discussion about the use of nuclear energy. The assumptions used to evaluate the occurrence of accidents in nuclear power plants can be doubted in view of the short interval of only 25 years between the catastrophes of Chernobyl and Fukushima. It is possible that the risks of nuclear power were systematically underestimated.
In March 2011, in response to the nuclear catastrophe in Fukushima, the German Bundestag decided to phase out nuclear power completely by 2022. (Reinberger, D. et. al., 2016; Filburn T, Bullard S, 2016)
Mathematically, an uncontrolled and prolonged release of radioactivity can occur in any reactor worldwide, with catastrophic consequences for humans and the environment. Individual traffic accidents generally do not have such a dimension—but in total they do.
According to statistics, the absolute frequency of dying in a road accident in 2018 was:
  • Approximately 3,000 annually in Germany
  • Approximately 40,000 annually in the USA
  • At least around 1,272,000 annually worldwide [4, 9, 10]
    $${\text{Global Traffic Mortality Rate}}_{2015} = { }\frac{{1,272,465{ }}}{7,313,015,000} = 17.4{*}10^{ - 5} { }\frac{{1{ }}}{{\text{a}}}$$
That means it is equal to 17.4 persons out of 100.000 who died in European road traffic in 2015 (World Health Organization, 2017).
$${\text{European Traffic Mortality Rate}}_{EU28, 2016} = { }\frac{{25,671{ }}}{508,326,680} = 5.05{*}10^{ - 5} { }\frac{{1{ }}}{{\text{a}}}{ }$$
This is equal to 5.05 persons out of 100.000 who died in European road traffic in 2016 (European Transport Safety Council, 2017).
In the year 2010, the EU renewed its road safety target to reduce road deaths by 50%. The reduction is based on 2010 until the year 2020. This corresponds to a reduction of 18.7% by 2016 compared with 31,595 people dead in 2010. It followed an earlier target set in 2001 to halve road deaths by 2010. The target was not quite reached because 55,092 people were killed in 2001. But at least the 42.7% achieved were not very far away. Figure 2.2 shows the average age expectancy of women and men compared to traffic mortality per 100,000 inhabitants.
Conversely, HIV/AIDS deaths increased from 300,000 in 1990 until 1.5 million in 2010. Non-communicable disease deaths rose by almost 8 million between 1990 and 2010. Cancer alone killed 8 million people in 2010, an increase of 38% over two decades. The number of fatality road injuries grew by 46% from 907,900 to 1,328,500 over 10 years but age-standardized road injury death rates only rose from 18.4 to 19.5 per 100 000.
ISO 26262 requires a significantly higher level of security with regard to the hardware failure rate compared to many other deadly risks accepted in reality. The overview in Fig. 2.3 addresses global mortality rates with exemplary causes of death for 1990 and 2010 and in addition the Automotive Safety Integrity Level “ASIL D” requirement with a hardware failure rate of less than 1 * 10–8 1/h.
Considering an agreement for reasonable safety and acceptable risk requires an international approach. These safety relevant challenges are undoubtedly connected to the current accepted “social values” that exist within our society.
In order to quantify automated vehicles’ potential safety benefits selected accident data collections will be presented and their respective pros and cons discussed.

2.4 Federal Road Traffic Accident Statistics in Germany

The Federal Statistical Office of Germany in Wiesbaden publishes monthly statistics on fatalities, injuries, and material damage in accordance with Section 1 of the StVUnfStatG (§1, German law on statistics of road traffic accidents). This data is provided by police stations, which are required to submit standardized records of reported accidents to state-level statistics offices (Statistisches Bundesamt, 2014).
Only extracts of this nationwide data is published online. Police investigations show the drivers’ driving errors and therefore a potential for increasing safety through automated driving (see Ch. 3.​3). All documented information is categorized into: type of road, age of all parties involved, and type of transport means. No specific documentation on vehicle details, injuries or accident reconstruction is available.

2.5 German In-Depth Accident Study (GIDAS)

Statistically reliable analysis of road-accident scenarios requires detailed data. In Germany, the GIDAS (German In-Depth Accident Study) database serves this purpose. It is recognized as one of the most comprehensive accident databases worldwide (Kramer, 2013; Zobel & Winkle, 2014). GIDAS has been financed by the Federal Highway Research Institute (BASt) since 1973 and The Research Association of Automotive Technology (FAT) since 1999. These days GIDAS prepare separate databases of approx. 2,000 accidents annually from the Hannover (since 1973) and Dresden survey areas (since 1999). Each documented accident contains up to 3,000 coded parameters: information on the environment (e.g. weather, road type, road condition), the situation (e.g. traffic, conflict, and manner of accident), the vehicles (type, safety equipment), personal details, injury data including accident reconstruction as well as photos (Winkle, Mönnich, Bakker & Kohsiek, 2009; Kramer, 2013; Zobel & Winkle, 2014; Schubert & Erbsmehl, 2013).
For further analyses, many cases are reconstructed with the PC-Crash simulation software by Dr. Steffan Datentechnik (Steffan H & Moser A 2016; Burg & Moser A, 2017; Castro, Becke & Nugel 2016). However, GIDAS data access is limited to car manufacturers and component suppliers taking part in the project. It contains only accidents resulting in personal injuries. Because only the Hannover and Dresden areas are surveyed, the findings have to be transferred to the whole of Germany via extrapolation (i.e. weighting and comparison with federal accident statistics, see Section 2.4).

2.6 Road Traffic Accident Statistics in the USA

The US National Highway Traffic Safety Administration (NHTSA) introduced the Fatality Analysis Reporting System (FARS) in 1975 and has documented fatal road accidents since then (National Highway Traffic Safety Administration NHTSA, 2014). In addition, the National Automotive Sample System – Crashworthiness Data System (Nass-CDS) has analyzed road accidents involving personal injury or severe damage using interdisciplinary teams, similarly to the German GIDAS since 1979 (O’day J, 1986).
However, unlike GIDAS, in-depth data collections for extended accident analysis in the USA offer no reliable accident reconstruction. For example, emergency braking functions cannot be assessed (Zobel & Winkle, 2014). The drop in US traffic accident fatalities since 1970 has been lower, at around 16%, than in Germany, at around 60% (Statistisches Bundesamt, 2014; National Highway Traffic Safety Administration NHTSA, 2014). This might be, among other factors, because of drowsiness due to longer distances driven in the US.

2.7 International Road Accident Data Collections

Various national official accident statistics have been merged into the International Road Traffic and Accident Database (IRTAD). Both fatalities as well as road accidents involving personal injury generally are included – they are distinguished by age, location and type of road use. The database is maintained by the Organization for Economic Cooperation and Development (OECD) in Paris. It contains data from: Argentina, Australia, Austria, Belgium, Canada, Chile, Czech Republic, Denmark, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Israel, Italy, Jamaica, Japan, Korea, Lithuania, Luxembourg, Morocco, Netherlands. New Zealand, Norway, Poland, Slovenia, Spain, Sweden, Switzerland, the UK and the USA (Amoros, 2009).
The data is publicly accessible online and is especially useful for comparing the data between member countries. It gives insight into the impact of different regulations and national/regional driving behavior (north versus south, for instance). However, detailed information on how the accident occurred is still missing. Besides, survey methods and data volumes differ in each country.
The Initiative of Global Harmonization of Accident Data (IGLAD) also aims to harmonize global in-depth traffic accident data. In 2010 European car manufacturers started IGLAD in order to improve road and vehicle safety. A standardized data scheme determines the accident data contained. This enables comparison between different countries. Initially IGLAD was funded by the European Automobile Manufacturers’ Association (ACEA). In the second phase, which started in 2014, the number of variables was extended to 93 regarding accidents, roads, participants, occupants and safety systems. Until now only limited data (between 50 and 200 cases, data years 2007–2012) from 11 countries (Australia, Austria, China, Czech Republic, France, Germany, India, Italy, Spain, Sweden and USA) have been accessible for research.

2.8 Accident Data Collections of Automobile Manufacturers

Continuous improvements in the effectiveness of vehicle safety systems currently in use remain a prime aim for car manufacturers and component suppliers. Therefore, interdisciplinary expert teams collect information on accidents involving current vehicles and carry out accident analysis at the scene together with hospitals and the police, thereby also fulfilling product monitoring obligations.
Moreover, manufacturers also analyze complex accident scenarios in order to comply with mandatory duty of care and observe potential product dangers that may arise during operation. According to Section 823 of the German code of civil law (BGB), a car manufacturer is liable for errors of its products’ damages resulting from intended or foreseeable use. A manufacturer is therefore obliged to collect and analyze information on vehicle use in conjunction with innovative systems. The more dangerous a product, the greater is the obligation to ensure and monitor a product’s safety during and after the development process (Matthaei et. al. 2015), (see Ch. 4).
As far back as 1969, Mercedes-Benz started investigating road accidents involving its Mercedes vehicles in cooperation with the interior ministry of Baden-Württemberg. Mercedes’ accident research had access to regular information over the telephone and insight into police accident files. Since at least the 1970s, other manufacturers like BMW have increasingly been studying and documenting accidents involving their own vehicles. Volkswagen (VW) has obtained information from the insurer’s association Haftpflicht-, Unfall-, Kraftversicherer-Verband (HUK-Verband) since the late 1960s and from the Hannover Medical School MHH (the predecessor of GIDAS) since 1985. VW accident research has been analyzing its own data since 1995 (Zobel R, Winkle T, 2014).
Detailed, interdisciplinary investigation of accidents by automotive manufacturers especially with the support of function developers involving the latest vehicle safety technology provide clear insights into the potential benefits of automated systems. However, few hundred cases annually which only involve a brand’s own vehicles are not statistically valid.

2.9 Accident Data of the German Insurance Association

The German Insurance Association (Gesamtverband der Deutschen Versicherungswirtschaft – GDV), focuses on damage incidences from motor claims where German insurers have to pay compensation based on their contracts. This information helps the GDV for example in grading insurance contracts, or in determining the potential savings through driver-assistance systems (Hummel, Kühn, Bende & Lang, 2011).
Insurers’ accident research has access to motor vehicle liability loss and collision damage waiver (CDW) cases reported to the GDV. Unfortunately, this data is not publicly available. No analysis takes place at the scene. The accidents are not recorded comprehensively. As soon as the question of liability to pay has been answered, the insurer’s interest in the particularities of a case ends. Therefore, there is very little detailed information on the accident cause of undisputed cases. In accidents with only one party and one vehicle involved (driving accidents), when a driver loses control of the vehicle, the cause of the accident remains uninvestigated. (Zobel R, Winkle T, 2014).

2.10 Accident Data Collections of Consumer Associations (ADAC)

In 2005, the German automobile club ADAC started researching accidents involving the ADAC technology center and the ADAC air rescue. Annually, information on around 2,500 serious accidents from rescue flights is collected in the ADAC database. Accident data is supplied from expert reports by motor vehicle assessors, the police, emergency physicians and fire departments (Unger, 2013).
The ADAC accident data lists and describes road accidents with seriously injured persons. They include aerial pictures including a vehicle’s final position as well as an in-depth medical diagnosis. Although the files are not publicly accessible, it is possible to access and evaluate the data individually. Unfortunately, the various persons investigating the accident do not compile their respective results for interdisciplinary reflection.

2.11 The Fundamentals of Accident Data Analysis

2.11.1 Level of Data Collection versus Number of Cases

The validity of accident data with regard to potential safety benefits depends to a large extent on the collection method. Usually interdisciplinary teams work together to carry out so-called in-depth surveys. Well-founded results can be achieved when function developers, accident analysis experts, doctors, and traffic psychologists are all involved in analyzing individual cases. But this depth of data collection tends to be restricted to a small number of cases, diminishing its statistical validity.
Evaluations from accident databases give an indication which measures are likely to increase traffic safety. A detailed accident analysis including a reconstruction of the accident encompasses a retrograde calculation of speeds based on traces of the accident, an investigation as to how the accident arose, a check for possible accident fraud, consideration to what extent it was avoidable, and biomechanics. An extensive knowledge of the given conditions and framework is necessary for an evaluation of future systems’ potential benefits based on these findings.
Currently, promising ideas on improving vehicle safety primarily come from a combination of accident analysis, existing experience and extensive research work. Accident research is one way to review the efficiency of existing automated vehicle functions and the need for further safety-enhancing functions. Below, basic terms of accident data evaluation will be explained.

2.11.2 The Validity of Areas of Action Compared to Areas of Efficiency

When comparing various accident data analyses, the way in which data is collected and the way it is processed have to be distinguished. Areas of action adopted under optimal conditions are often confused with areas of efficiency under real conditions.
An area of action comprises the accidents which a system can influence. The area of action may vary according to how precisely a system’s specification is defined. As a result, this is an initial estimate of the maximum potential of the automation level in question. On the other hand, the actual resulting efficiency of a function is generally significantly lower. Efficiency is defined as the effect that a specified system has in practice. It is either proven by occurring accidents (a posteriori) or predicted by simulations (a priori) (Winkle T, et. al., 2009a).
Determining an area of efficiency therefore requires precise knowledge of two factors:
  • the system specification with its corresponding function limits
  • the driver’s behavior
The level of efficiency describes a function’s relative efficiency as a percentage and relates to the unspecified term of the area of action (Schittenhelm et. al. 2008):
$$\user2{degree}\;\user2{of}\;\user2{efficiency} = \frac{{\user2{area}\;\user2{of}\;\user2{efficiency}}}{{\user2{area}\;\user2{of}\;\user2{action}}} = \user2{x}\;[\%]$$

2.11.3 Potential Safety Benefits Depending on Automation Levels and Degree of Efficiency

Some analyses of potential safety impacts examine the maximum assumed area of action described above by using accident databases. In contrast, analyzing the degree of efficiency comes closer to reality by evaluating an area of efficiency for its actual benefit (Schittenhelm et. al. 2008). However, the resulting safety benefits of automated vehicles can only be established after all risks have been factored in. The benefit corresponds with the reduction of accident frequency and severity. New risks exist since as yet non-existent accidents may occur with increasing automation.
The theory of inventive problem solving (TRIZ) defines the requirements of an ideal machine, using the formula of an ideal final result with an unlimited benefit, while incurring no costs or damages (Hummel T, Kühn, Bende & Lang, 2011):
$$\user2{ideal}\, \user2{final}\, \user2{result} = \frac{{\sum {\varvec{benefit}}}}{{\left( {\sum \user2{costs } + \user2{ }\sum {\varvec{damages}}} \right)\user2{ }}} = \frac{\infty }{{\left( {0\user2{ } + \user2{ }0} \right)}} = \infty$$
On the one hand, the safety benefit of connected automated vehicles increases in accordance with the degree of efficiency (proof by accident data analysis and knowledge of functions). On the other hand, the risks may rise in line with an increase in automation (“Driver” versus “Robot”). These in turn lessen the actual safety benefit (see Fig. 2.4). To minimize those potential risks, manufacturers carry out risk management (see Ch. 2) using accident data.

2.12 Significance of Possible Predictions Based on Accident Data

Using exemplary cases, the following meta-analysis shows what conclusions can or can’t be drawn about potential benefits based on various accident data. Since there have been no analyses yet of highly and fully automated vehicles, we will look at systems without automation (“driver only”/”no automation”) or with low levels of automation concerning the main driving task (“assisted”/”partially automated”) first and divide them into a posteriori and a priori analyses.
Section 1.​4.​1 contains examples of a posteriori statements on accident data. In the definition used here, figures “gained from experience” (Duden, 2014) can be interpreted immediately. In contrast, assumptions “obtained by logical reasoning” (Duden, 2014) must be made in order to assess the potential benefits of future levels of automation when using the a-priori- forecasts defined in Section 1.​4.​2., which are based on accident- data collections.

2.12.1 A Posteriori Analyses of Accident Data for “Driver Only”/“No Automation”

Past and present a posteriori analyses of accident data collections involving conventionally (human-) driven vehicles provide insights into accident black spots and changes in real-life traffic accidents. This “driver-only”/“no-automation” category means a lack of warnings and interventions in longitudinal and lateral guidance by environmental sensors.
The change in the number of accident fatalities serves as a first example. The second example is the positive impact of Electric Stability Control, or ESC (see Ch. Traffic Statistics: Accident Fatalities Versus Registered Motor Vehicles

The rate of traffic accident fatalities per registered vehicle, taken from data of the German Federal Statistics Office shows that death rates have been dropping in Germany since 1970 when 21,332 people died in car accidents (Statistisches Bundesamt, 2020). Since then, the numbers of injuries and fatalities in road accidents have been reduced considerably in Western countries due to measures in road building, legislation, the rescue chain, emergency medicine, and passive and active vehicle safety. These findings are based on large-scale worldwide collected surveys and analyses of road accidents. They are affected by various orientations, different amount of data and based on investigations of varying depth.
These accident statistics show that, while the number of registered vehicles increased, the number of traffic fatalities dropped from over 21,000 in 1970 to almost 3,000 annually. This was due to various legislative, medical, technological, and infrastructural measures (see Fig. 2.5). Because of the overlapping of all these actions, it is difficult to single out and calculate the effectiveness of any individual measure. Studies on the Effect of “Driver-Only”/“No-Automation” Systems

Introduced in 1995, Electronic Stability Control (ESC) is a technical evolution of the electronic antilock braking system, which was largely marketed from 1978 with the legally protected term ABS. It uses ABS’s wheel speed sensors in conjunction with additional sensors for steering wheel angle, yaw rate, and lateral acceleration. Using this information, ESC can stabilize the vehicle in case of a recognized skid through braking individual wheels independently of each other. With this braking intervention, a lateral collision can be converted into a less vulnerable frontal crash. In 2001, Daimler accident research posited that 21% of skidding accidents led to injuries and 43% to fatalities (Daimler AG Communications, 2011). At that time, the findings of accident research experts investigating individual accidents on behalf of car manufacturers diverged greatly. Later forecasts of potential benefits based on a larger number of cases also differ. Areas of action from the year 2000, for example, show a positive impact of up to 67% for severe skidding accidents (Bengler et. al. 2014). Other studies stated that, second only to the introduction of safety belts as a passive safety system, ESC provides the most effective gain in safety in the “driver-only” category (Zobel et. al. 2010). The proportion of accidents due to driver error and skidding, for instance, decreased after the introduction of ESC as a standard in all Mercedes-Benz cars from about 2.8 involved vehicles (per 1000 registered in Germany) in 1998/1999 to 2.21 involved in 2000/2001. ESC’s high effectiveness has also been proven in other brands such as Volkswagen, where statistics show lower accident frequency as well as prevention of critical accident types (Langwieder, Gwehenberger, Hummel, 2003).
In summary, safety benefits have already been established for safety-enhancing “driver-only” functions with quick market penetration depending on suppositions and various data sources. Especially for ESC, the scientific evidence for an increase in safety is well-founded.

2.12.2 A Priori Predictions for Assisted and Partially Automated Driving

A priori predictions depend on hypotheses and inferences. For example, assisted and partially automated driving functions can keep the driver from imminent danger via acoustic, optic or haptic warnings as well as short braking or steering interventions with a warning character. However, the danger can only be successfully averted if the driver reacts in time and appropriately to the traffic situation.
From a technical viewpoint, these advanced levels of automation, which possess a greater degree of extended computer and sensor technology for environmental perception, result in increasingly capable assistance systems. Some currently available safety-enhancing driver assistance systems warn the driver when there is recognized danger in parallel or crossing traffic. These include collision warning systems such as EBA – Electronic Brake Assist, ACC with FCWS – Adaptive Cruise Control with Forward Collision Warning System, LKA – Lane Keep Assist, LDW – Lane Departure Warning, NV – Night Vision or intersection assistance. Other systems, such as Electronic Brake Assist (EBA) or Autonomous Emergency Brake (AEB), intervene in the longitudinal and lateral vehicle dynamics (see Fig. 2.5). Study on the Potential of Lane Departure Warning

Using the example of a Lane Departure Warning (LDW) system (Hörauf, Buschardt, Donner, Graab & Winkle, 2006), road accidents were analyzed by doctors, psychologists and development engineers in a cooperative approach in 2006. The results, which were obtained with the participation of the author of this book, a function developer, and a psychologist, were achieved through interdisciplinary research of a car manufacturer, a university hospital, and the police, with support from the Bavarian Ministry of the Interior, Building and Transport (BStMI).
Such interdisciplinary analyses of accident causes and consequences are based on technical, medical, and psychological examinations by experts from each field, which are then integrated collectively. These days, driving-related psychological data is collected more frequently in order to analyze a road accident. With the help of standardized interviews, the collision experience is recorded and evaluated from the driver’s viewpoint. The purely technical reconstruction of the accident is now supplemented by a psychological perspective.
Taking the example of Lane Departure Warning, it was explained to all professional teams involved what system design specifications had to be met. The selected accidents were filtered further through specific focused questions from the technological development. This kind of procedure provides insight into what kind of and how many accidents could be avoided through systems currently under development. To achieve this, knowledge of the system’s specific technical limits is indispensable. A further outcome may be recommendations for additional functional system enhancements (Hörauf, Buschardt, Donner, Graab & Winkle, 2006).
Therefore, these detailed accident analyses prove the value of comprehensive accident data collection. Experts on technology, medicine, and psychology worked together closely for this study. This interdisciplinary approach produces a large number of new references regarding accident scenes, vehicle details, injury patterns, parties involved in an accident and witness statements. This additional information gives insight into active steering corrections, interventions of the brakes and reactions immediately prior to a collision, since human errors such as inattentiveness, distraction or fatigue are the main causes for lane departure. The various perspectives from which an interdisciplinary team looks at the accident can make computer-aided reconstruction and simulation of an incident highly realistic. However, to achieve representative results, these analyses need to be validated by larger accident data collections. Interdisciplinary Degree of Efficiency Analysis Based on Current River Assistance Systems

Now that the advantages of interdisciplinary analysis had been proven through the above-mentioned study on the effectiveness of Lane Departure Warning, a further interdisciplinary analysis of the degree of efficiency was conducted four years later. The objective was a comparison of available safety-enhancing driver assistance systems. This project was based on a sample of reconstructed accidents (n = 100). Therefore, an interdisciplinary accident data evaluation was carried out by the author in cooperation with a psychologist and in close consultation with the respective function developers. The study analyzed the effectiveness of various driver assistance systems in avoiding accidents with regard to the accident situation (Chiellino, Winkle, Graab, Ernstberger, Donner, Nerlich, 2010). In early 2010, the range of systems available included Night Vision, Lane Departure Warning, Lane Change Assistant and Adaptive Cruise Control. To calculate the degree of efficiency, accident research data was weighted according to accident statistics for Bavaria. An accident scene was reconstructed for each real-life accident, and the accident cause in terms of human-machine interaction was assessed. This was done according to the human-machine interactions as described in the ADAS Code of Practice definition for the development of Advanced Driver Assistance Systems (ADAS) with active longitudinal and lateral guidance (Donner, Winkle, Walz, Schwarz, 2007). After a six-year involvement (Becker, Schollinski, Schwarz & Winkle, 2003; Becker et. al. 2003), the European Automobile Manufacturers’ Association (Association des Constructeurs Européens d’Automobiles—ACEA) published the results in 2009 (Knapp, Neumann, Brockmann, Walz & Winkle, 2009). The potential for preventing accidents was judged to be positive only if every development expert for the relevant system saw its benefits. The results yielded that the examined systems were able to contribute significantly to diminishing the severity of accidents.
The study’s prognosis is that the investigated driver assistance systems would prevent a substantial number of accidents. A 27% decrease in the total number of injured persons was predicted, which means that the number of people injured would fall from 126 drivers and 49 passengers (as in the actual data) to 94 and 33, respectively. One must keep in mind that the premise for these results is optimal reactions regarding human-machine interactions. Further studies with test persons are necessary before drawing final conclusions. Moreover, 100% distribution of the investigated systems, operating without errors within the system limits, would need to be ensured.
The study used an injury grading system which was based on the Abbreviated Injury Scale (AIS) (Association for the Advancement of Automotive Medicine, 2005), as also applied in ISO 26262 for functional safety (International Organization for Standardization, ISO 26262–3, 2018). The AIS codes every injury with a value between 1 (light injuries) and 6 (extremely critical or fatal injuries). Thus, the most severe injury of all the injuries one person has contracted is defined as MAIS (Maximum AIS). An uninjured person is classified as MAIS 0.
Looking closely at accident causes revealed that over 60% of them involved information errors, i.e. failures regarding information access and information reception. Therefore, the correspondingly high effectiveness of warning assistance systems is hardly surprising (Chiellino, Winkle, Graab, Ernstberger, Donner & Nerlich, 2010).
In summary, this interdisciplinary study compared currently available driver assistance, with all respective developers being involved in the analysis. Each developer contributed their knowledge of the specific relevant function parameters of their system, thus ensuring more accurate assessment of potential gains in safety. It has to be born in mind that the sample of 100 cases in the study, weighted with representative accident data from Bavaria, is too small to yield statistically reliable statements. However, they show a tendency in which cases these driver assistance systems contribute significantly to road safety.
It is noteworthy that there are further possibilities for gaining statistical evidence regarding the predicted safety benefits of braking assistance and automatic emergency braking functions. Moreover, simulations using software-based accident reconstructions are immensely useful for assessing the forecast safety gains (Busch, 2005). GIDAS Database Analysis for Potential Safety Benefits of Connected Vehicles

Using a larger data volume, the following analysis of the German In-Depth Accident Study (GIDAS) database demonstrates the variety and complexity of several assumptions. In cooperation with a team of experts, the author conducted this analysis in 2009 as part of the Safe and Intelligent Mobility – Test Field Germany (Sichere Intelligente Mobilität: Testfeld Deutschland – simTD) research project with a more significant sample. The aim was an assessment of the potential benefit of future safety-relevant automobile communications systems. The analysis included functions for connected systems with an immediate safety impact on road traffic. The relevant data was gleaned from 13,821 accidents involving personal injury, which had been documented by GIDAS between 2001 and 2008 in the areas of Hannover, Dresden, and their surroundings (Winkle, Mönnich, Bakker & Kohsiek, 2009; Schubert & Erbsmehl, 2013). In order to extrapolate this for the whole country, the data obtained from the statistical sampling scheme was weighted with the help of accident statistics from the German Federal Statistical Office. These official statistics list all accidents registered in Germany in one calendar year which involve personal injury. For example, there were 335,845 road accidents involving personal injury in 2007 (Statistisches Bundesamt, 2014).
In several consultations with the simTD function developers and accident experts from BMW, Audi, Daimler, Bosch and Volkswagen, the precise variables needed for the analysis were agreed on. The project participants decided to start with the analysis of 13 safety-related warning functions. They made a joint decision to consider relevant vehicles such as cars, trucks, agricultural tractors, buses, rail vehicles (including city railways and trams, but no state railway trains) and motorbikes (motorized two-wheelers, three-wheelers, quad bikes from 125 cc) during several workshops. After this, the areas of action using the extensive GIDAS data were determined. Initially this selection was made by using the variables from all accidents relevant to each system as they related to the whole of the accident occurrence. The result was that, ranging from 0.2% to 24.9%, the areas of action for each separately examined function varied greatly. Areas of action can therefore give a fairly certain estimate only of the maximum effectiveness which cannot be exceeded. It should also be noted that due to overlapping functions individual areas of action cannot simply be added up (see Annex Fig. A.18).
In order to analyze degree of efficiency, three assumed function types (electronic brake light, cross traffic assist, traffic sign assist for stop signs) were selected from the GIDAS area of action analysis mentioned above. The corresponding degrees of efficiency were taken from a reduced sample of driving simulator investigations.
For instance, in accidents where cross traffic assists helped the driver to avoid them (see Klanner, 2008), there was a considerable range, from 9.9% to 73.3%. This was due to both different driver reaction times and varying braking intensity after warnings. Thus, three likely reaction times (0.54, 0.72 and 1.06 seconds) and the probabilities for the occurrence of each one were determined. In addition, weak braking of 50% of maximum braking pressure was assumed for unsuccessful reactions and 100% for successful reactions (Winkle, Mönnich, Bakker & Kohsiek, 2009; Schubert & Erbsmehl, 2013).
The objective of this elaborate approach to analyzing degrees of efficiency was to determine and evaluate the potential of future, connected, safety-enhancing driver assistance functions with statistical relevance. However, the wide range of up to 70% which was found decreases the validity and therefore only yields tendencies and outlooks regarding accidents avoided. This vast scattering is a result of the sensitivity of the parameters depicted above and the warning algorithm in question, as in practice drivers’ reaction times and braking intensities vary greatly.

2.12.3 Potential Safety Benefits and Test Scenarios for Development of Highly and Fully Automated Driving GIDAS Database Expert Estimates Until 2070

From a technical viewpoint, under favorable conditions current automated vehicles can already autonomously carry out many driving tasks in moving traffic. Whereas driver assistance systems merely support the driver, advanced systems like highly and fully automated driving temporarily or permanently take on the task of driving.
Highly and in particular fully automated driving is engineered to approach “Vision Zero”: traveling as accident-free as possible. Roads and means of sustainable transportation ought to be planned and constructed in such a way that there are no traffic accident fatalities or severely injured victims. The accident-free vision originated in occupational safety and was first applied to road traffic in the 1990s in Sweden. The EU backed projects for connected automated vehicles like the “Highly Automated VEhicles for intelligent transport” (HAVEit) research project, which it sponsored with 17 million Euros. Car manufacturers such as Daimler, BMW and Volkswagen/Audi are also working on the vision of accident-free driving. Thomas Weber, former Board of Management member of Daimler AG for research and development, asserts in an interview:
“Unser Weg zum unfallfreien Fahren treibt uns an, die Mobilität auch in Zukunft für alle Verkehrsteilnehmer so sicher wie möglich zu gestalten.” (Daimler AG Communications, 2011)
(Our ‘path to accident-free driving’ also drives us to design mobility as safely as possible for all road users in the future)
In the first decade of this century, the number of road accidents with a car as the main cause and resulting in personal injury fell in Germany from 266,885 in 2001 to 198,175 in 2010. At 68.7%, cars are still the main cause of road accidents according to the Federal Statistical Office (2010). The accident types can be broken down into the following main categories: Turning at/crossing intersections (58,725), parallel traffic (44,812), turning (33,649) and 30,737 dynamic accidents (Statistisches Bundesamt, 2014) (see Fig. 2.6).
To date, we don’t have empirical proof of the cumulative safety increases of fully automated driving functions. Daimler compiled one of the first comprehensive forecasts in vehicle safety and accident research. It investigated the potential of automated vehicles regarding accident prevention based on assumed deployment and market penetration scenarios. For these they relied on expert estimates, third-party forecasts and GIDAS data. The forecast provides an initial rough estimate and is based on a total of 198,175 preventable accidents caused by cars in 2010 (see Fig. 2.6).
The assumptions include changes within each type of accident (parallel traffic, stationary traffic, pedestrians, turning at/crossing intersections, turning, dynamic accidents). For instance, the pie charts show that accidents involving a car in parallel traffic or losing control will decline by around 15% by 2060 with increasing automation, while accidents when turning at or crossing intersections will proportionately rise by around 10% (Unselt, Schöneburg & Bakker, 2013).
According to Daimler’s estimates for increasing automation, an overall decrease of 10% of accidents was achieved by 2020. In the following decades, reductions of 19% could be achieved by 2030, of 23% by 2040, of 50% by 2050, of 71% by 2060 and almost complete prevention by 2070 (Unselt, Schöneburg & Bakker, 2013). The forecast thus predicts that in 2070 an autonomous car will cause nearly no accidents, but may be at risk of being involved in serious collisions. It can safely be assumed that an automated car will be able to prevent some collisions that another vehicle would have caused. However, it should be noted that this study does not include accidents caused by other road users. Potential technical failures (see Fig. 2.9) are also outside its scope. Furthermore, the data stemming from the German Federal Statistical Office, and above all the validity of GIDAS, mainly relies on crash and post-crash statements by injured people (see Schubert, Erbsmehl & Hannawald, 2012). World-Wide Accident Data Evaluation for Relevant Traffic Test Scenarios

To obtain a comprehensive evaluation of highly and fully automated vehicles’ active safety in a development lifecycle (see Fig. 2.7), the author recommends incorporating findings from accident data collections around the world as well as analysis of incidents not resulting in injuries, near collisions, traffic simulations and weather data.
Therefore, a first-time area-wide study based on all police reports has been carried out. The findings can be supplemented with information from hospitals, insurance companies and human behavior models. Once all relevant factors that can lead to a collision are known, virtual simulations based on quantitative and trained neural (e.g. AI) models can be performed. Possible system responses would be classified as true positive/true negative and false positive/false negative. The evaluation of automated safety functions should consider all possible system responses (Helmer, 2015).
The aim is to combine all known accidents by using geographically defined road accident data in conjunction with high-definition geographic digital mapping data (e.g. Google Maps, TomTom, Nokia HERE, OpenStreetMap) as well as traffic flow data from various sources (e.g. vehicles, cell phones, road traffic devices). For example,
SAFE ROAD MAPS (http://​www.​saferoadmaps.​org) provides localized collision data in the US. The UK publishes similar details on www.​data.​gov.​uk; these in turn are integrated into the UK Road Accident Map. German regional accident data can be obtained from police IT applications. These depend on the federal state and include the Geographical Positioning, Analysis, Representation and Information System (Geografisches Lage-, Analyse-, Darstellungs und Informationssystem—GLADIS), the Road Accident Location Map and Analysis Network (Verkehrs-Unfall-Lage-Karten und Analyse-Netzwerk – VULKAN), the Geographical Police Information System for Road accidents (Geografisches Polizeiliches Informationssystem für Verkehrsunfälle – GEOPOLIS V), the Brandenburg Expert System for the Analysis and Documentation of Accident-Heavy Route Sections (Brandenburgisches Expertensystem für die Analyse und Dokumentation von unfallauffälligen Streckenabschnitten – BASTa) or the widely used Topographical Electronic Accident Type Map (Elektronische Unfalltypensteckkarte – EUSka) (Dick, 2011).
Currently, however, there is still a lack of precise specifications for OEM (Original Equipment Manufacturers) mass production solutions that are ready for market launch as well as reliable descriptions of the functional limits of highly and fully automated vehicles. Thus, to date forecasts of potential safety benefits rely heavily on numerous assumptions. Reliable data on market launch and penetration is also not available. Hence current predictions of potential safety benefits, which are solely based on accident data, have limited validity. It is therefore advisable to link in-depth accident data collections (e.g. GIDAS) with all available global accident data collections and analyses, traffic simulations, vehicle operation data and related weather information.
The learning curve in figure 2.8 demonstrates the increasing amount of available real-world data of automated vehicle functions before and after market launch. For the identification of relevant critical scenarios, the author recommends regular monitoring and analysis of all available data of automated functions (see also Annex A.17). These supply knowledge for sensor simulation, image classifications and decision strategies regarding future connected automated vehicles.

2.13 Potential Safety Benefits / Risks and Impacts on Testing

2.13.1 Human Error versus Technical Failure in Full Automation

Human performance in driving can be increased. The metaphorical example of the interaction of horse and rider shows that in the cooperative guidance of movement (see H-mode) redundant cooperation partners complement each other in their abilities with regard to perception and action, such as experience or tiring situations (Bengler K, Flemisch F 2011). First of all, a fully automated vehicle must reach this safety level. Only fault-free fully automated vehicles will be able to come close to “Vision Zero”. On one hand we have to consider the human error in the causes of accidents on the other hand driving experience should not be underestimated. Machines can only handle driving situations that have been programmed. Beyond that, fully automated self-driving cars are restricted due to physical or technical limits.
Based on the GIDAS accident database, the left-hand side of Fig. 2.9 shows the statistical distribution of accident causes. At 93.5%, “human error” is the main reason for road accidents. Compared to that, the impact of unfavorable driving conditions or the environment (for example road surface quality or the weather) is at 4.6% quite low, with technical failure being even lower at 0.7% (Volkswagen/German In-Depth Accident Study, 2010).
Naturally, the possibility of accidents due to driver error is eliminated completely during fully automated driving sections. The “technical failure” category could therefore increase proportionally, with the added technical risks of full automation. As a consequence, the public can be expected to give it more attention (see Fig. 2.9).
In the future, further evaluation and overcoming human error processes in real-life traffic situations (supplemented by global relevant test scenarios which are based on comprehensively linked up and geographically defined accident-, traffic flow- weather- and vehicle operation data collections) will facilitate virtual traffic simulations for safe development, tests and validation of automated cars (Kompass, Helmer, Wang & Kates, 2015).

2.13.2 Potential Safety Benefits – Human and Machine Performance

Car traffic safety today relies mostly on human skills and their support by safety-enhancing systems. Fully automated vehicles will depend only on machine performance. According to the level of automation, humans’ perceptions, experience, judgment and capacity to react will be replaced by technical systems. The potential safety benefits as well as the risks of increasingly automated driving can be attributed to the various strengths and weaknesses of both humans and machines.
For instance, machines can neither react appropriately to unknown situations nor interpret the movements of children (see Dietmayer et. al., 2015; Dietmayer, 2016). On the other hand, people can be inattentive, misjudge speeds and distances and have a more restricted field of vision than machines (Knapp, Neumann, Brockmann, Walz & Winkle, 2009).

2.13.3 Artificial Intelligence versus Human Perception Limits and Consequence

To demonstrate the limited machine perception and Artificial Intelligence in comparison with human perception, a simplified model of current sensor technologies in use is described below. A vehicle requires sensors in order to collect information about its environment. Sensors can be classified according to their physical measuring principle. Cars mainly use radar, lidar, ultrasound sensors, near and far infrared, and cameras (see Maurer, 2000; Siedersberger, 2003).
The top and center image of figure 2.10 illustrate simplified and color-coded measuring principles that lead to limited machine perception. The bottom image superimposes all the above-named measurements onto what human drivers can see in difficult light- and weather conditions (sun, backlight, wet road surface, spray/splashing water, icing/contamination of windshield/sensors, road markings only partially visible). A closer look shows that the radar reflection point (blue) on the left is a false detection, which has been caused by a reflection in the other lane (see Becker et. al. 2004; Donner et. al. 2004).
Figure 2.10 illustrates that the outcome of machine perception and interpretation of complex traffic situations continues to present development engineers with considerable technical challenges. These include detecting static and dynamic objects, physically measuring them as accurately as possible, and allocating with the correct semantic meaning to the detected objects (see Dietmayer, 2016).
Difficult light- and weather conditions challenge human and machine perception in real traffic situations. For this purpose, area-covering accident data analyses are able to indicate temporally and geographically related accident black spots. To analyze scenarios with reduced visibility due to fog, rain, snow, darkness and glare from sun or headlights, the author carried out a first-of-its-kind area-covering accident study in cooperation with Christian Erbsmehl from Fraunhofer Institute for Transportation and Infrastructure Systems IVI in Dresden (see Ch. 3).

2.13.4 Human Error versus Artificial Intelligence Incertitudes

Advancing vehicle automation of the main driver tasks result in new research questions. Attentive and vigilant drivers have substantial skills to deescalate dangerous traffic situations. Human’s capabilities provide significant input for traffic safety today. Differentiated potential benefit estimates would need to compare the performance of humans and machines. Especially takeover situations between driver and machine involve new challenges for design and validation of human-machine interaction. Initial tests at the chair of Ergonomics at Technical University of Munich (TUM) demonstrate relevant ergonomic design requirements which will be continued (Bengler, 2015).
Fundamental correlations between automation and human performance can be evaluated by many methods. It is possible to identify the probability of a road accident by the use of a fault tree. Amongst others the probability includes human failure, inappropriate behavior and the existence of a conflicting object (Reichart, 2000). The choice of actions to avoid a collision is greater, if the potential road accident is less imminent.
The evaluation of driver behavior requires observations for a longer period. Regarding human failures analyzing the perception process chain provides in-depth knowledge. Such analyses draw on evaluations of psychological data from road accidents (Gründl, 2006). In terms of interdisciplinary accident analysis, an error classification of five categories has approved by practical experience in accident research. This five-steps method is a further development of ACASS (Accident Causation Analysis with Seven Steps). It was developed jointly with GIDAS along the lines of the seven-step principle from Jens Rasmussen, former system safety and human factors Professor in Denmark, a highly influential expert within the field of safety science, human error, risk management and accident research (Rasmussen,1982). Using the five-steps method it is possible to identify human errors, define the time during the perception process from accessing the information to operation, and to evaluate the particular type of error (see Fig. 2.11). The associated questions concern: Information access (was the relevant information of the traffic-situation objectively accessible to the driver? Was the field of vision clear?), information reception (did the driver observe the traffic situation properly and perceive/detect the relevant information subjectively?), data processing (did the driver correctly interpret the traffic situation according to the available information?), objective target (did the driver decide appropriate to the traffic situation?), and operation (did the driver carry out his or her decision into operation properly?).
Using this classification, the accident analysis shows that the predominant sources of human error lie in information access and reception (see Fig. 2.11); (Chiellino, Winkle, Graab, Ernstberger, Donner & Nerlich, 2010; Weber, Ernstberger, Donner & Kiss, 2014).
Regarding accident statistics with reference to human driving errors as the stated cause of accidents, the proportion of driving failures is quantified with: 93.5% (source GIDAS). In addition, probabilities are indicated with: evasive stress action to mitigate imminent crash is indicated by p = 0,1 … 1; evasive action with sufficient time gap is indicated by p = 10–1 … 10–2 and trained lane keeping is indicated by p = 10–4 … 10–5 (Bubb H, Bengler K, Grünen R-E, Vollrath M, 2015; see also Fig. 4.​11).
For Artificial Intelligence perception, Klaus Dietmayer, Professor in Ulm at the Institute of Measurement, Control, and Microtechnology, Expert for Information fusion, Classification, Multi-Object Tracking, Signal processing and Identification (see Dietmayer, 2016) names three essential domains of incertitudes corresponding to human information access as well as data processing. These three are: firstly state-, secondly existence-, and thirdly class uncertainty. All three have a direct impact on machine performance. If the uncertainties in these areas increase beyond a yet to be defined “tolerable limit”, errors in the automatic vehicle guidance can be expected. In terms of making forecasts, only an indication of trends is currently possible.
“While the currently known methods for estimating state and existence uncertainties do not enable a current estimation of the capability of the machine perception, in principle it is not possible to predict degeneration in the capability of individual sensors or even a failure of components.” (see Dietmayer, 2016)

2.13.5 Potential Safety Benefits of Fully Automated Vehicles in Inevitable Incidents

When analyzing the potential safety benefits of fully automated vehicles, it is also important to consider persistent risks in the area of complex traffic situations and today’s known inevitable incidents. These include accidents at poorly visible and unclear intersections or behind visual obstructions. In a study of individual cases as part of a doctoral thesis at the University of Regensburg, visual obstruction was identified as a contributory cause in 19% of all cases (Gründl, 2006). Examples include trees, bushes, hedges, and high grass. Obstructions for instance may also be the cause of an accident if a child running out suddenly and unexpectedly in front of a car from between parked vehicles or a yard entrance.
This especially includes errors in the sequences of the perception process, in the accessing and reaches its limits.
Due to the large number of possible and non-predictable events, especially the reactive actions of other road users, the uncertainties increase so strongly after around 2 s to 3 s that reliable trajectory planning is no longer possible on this basis.” (see Dietmayer, 2016)
Therefore experience-based, internationally valid guidelines with virtual simulation methods for verification of automated vehicles and final testing of the overall system limits in a real environment are recommended. This includes interaction tests with control algorithms and performance verification of real sensors in real traffic situations, particularly at the time just before a collision (Schöner, Hurich, Luther & Herrtwich, 2011; Schöner, 2015).

2.14 Conclusion and Outlook

The findings from road accident research confirm: human failure is the main cause of road accidents. This especially includes errors in the sequences of the perception process, in the accessing and reception of information.
In order to estimate the potential safety benefits of highly and fully automated vehicles from accident data, a sophisticated comparison of the overall performance of humans and machines is required (see Annex Fig. A.16). This, however, will only be possible when precise knowledge is available concerning the functional characteristics and technical limits of developments planned for mass production.
Statistically verified expert assessments have already proven the potential benefits of future safety-supporting vehicle and driver assistance systems. Even before development begins, the developer can assess potential benefits. Additionally, by analyzing and evaluating traffic accidents after market launch, car manufacturers can fulfill their product monitoring obligations.
Overall, the results of road accident analysis today verifiably show that automating driving tasks from the “driver only”, “assisted”, up to “partially automated” driving categories are key technologies in contributing to minimizing the consequences of human failure.
Forecasts for highly and fully automated vehicles, generated using traffic accident data, only give results based on numerous assumptions. A forecast of fully automated vehicles’ potential safety benefits came from a first Daimler accident research appraisal that is based on several expert assumptions. According to Daimler’s estimates, practically complete elimination of accidents is possible by 2070 – assuming successful market penetration. However, according to the definition given in the publication only accidents triggered by cars were looked at, and no consideration was given to physical limits and potential technical defects. This appraisal is thus based on some assumptions still to be refined and validated more detailed in the future.
Above all, the possible technical potential (for example, unknown advances in Artificial Intelligence for machine perception) limits an accurate forecast. In particular, development engineers are faced with considerable technical challenges when perceiving and interpreting complex traffic situations. Furthermore, human performance is often underestimated. According to findings from traffic accident analyses, assistance and partially automated systems are generally capable to compensate weaknesses of human capabilities. They can increase safety in routine human driving situations with supervision, warnings and lateral or longitudinal support. On the other hand, to further reduce the number of traffic accidents, driverless vehicles must at least match the driving skills of an attentive human driver (supported by assistance and partly automated systems) before series development can be considered. Only when these technical barriers have been overcome, can a large-scale rollout of marketable fully automated vehicles be expected.
Until then (as an alternative measure for the assessment of potential safety benefits) assumptions of an assumed technical system configuration and system design have to be made without knowing the system limits or failure rates.
In summary, the following issues limit the validity of the potential safety benefit forecasts from “driver-only” to fully automated vehicles and will have impact for testing:
  • Fully automated vehicles’ degree of efficiency cannot be precisely quantified at present, as numerous technical and market-specific factors are still not known in detail. The evaluation of automated safety functions has to consider all possible system responses: True positive (or negative) and false positive (or negative).
  • The potential safety benefits stated four levels of automation so far (from driver-only to advanced functionalities) and should be judged and used with care, depending on the data used. The validity and forecasting reliability of the data material both depend on the selection and evaluation of available parameters.
  • Various approaches to evaluating potential benefits are to be compared with each other under expert consideration. Areas of action show the ideal maximum of possible preventable road accidents. In contrast to this is the actual identifiable efficiency, which is considerably lower.
  • The validity of evaluation methods can vary greatly: In addition to experienced accident investigators, it is recommended to involve medics, psychologists and development experts for automated functions in the analyses. Such multi-layered background information allows him or her to get a complete overview of a complex accident incident and reconstruct or analyze it more precisely than a colleague without this detailed knowledge.
  • There are often many overlapping areas of action within and between analyses of potential benefits reducing the overall area of action.
  • To obtain further findings for the development and design of safe automated vehicles (see Ch. 2), existing in-depth surveys of severe road accidents involving personal injury (e.g. GIDAS) should be combined with available area-covering accident collision data, digital geographic mappings, weather data and virtual traffic simulations (see Ch. 3).
  • Starting from the level highly automated and beyond, persons involved in an accident have – temporarily at least—no responsibility for the controllability of the vehicle. Measures to reduce risks and guarantee the functional safety of electrical and/or electronic systems are thus of prime importance.
  • It may be assumed that individual accident scenarios may still arise as a result of increased degrees of automation, right up to full automation in spite of rule-consistent way of driving. This applies, for instance, to physical driving limits or time-critical situations, such as a child running suddenly in front of a vehicle.
  • Area-wide accident analyses provide relevant scenarios for testing and verification of automated vehicles including virtual simulation methods, but final testing of the overall system limits in a real environment will not be completely eliminated.
Even if the technology of driverless cars never reaches 100% perfection, and a few as yet unknown accident scenarios arise as a result, the vision of area-covering driverless vehicle use in road traffic appears to promise a socially desirable benefit. Research activities that make use of interdisciplinary experts working on vehicle automation should therefore be promoted and strengthened. It is recommended to combine in-depth accident data with all worldwide geographically defined accident data collections, related weather- traffic flow and vehicle operation data information considering data protection measures. This will lead to actual safety benefits and statistically relevant scenarios for development including validation or testing of automated driving pertaining to machine versus human perception.
Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://​creativecommons.​org/​licenses/​by/​4.​0/​), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.
Findings from Traffic Accident Analysis
verfasst von
Thomas Winkle

    Premium Partner