Search

New research challenges and low-cost technological solutions drive the motivation to record behavior in multivariate ways using high temporal resolution. Making such data accessible and usable is more complicated than it may seem. Methods like ECG, EEG, and eye tracking can produce very large amounts of data in a short time. Further, context data to explain the observed behavior must be recorded as well. E.g., in a field study using an instrumented research vehicle, the position of the vehicle and the distance to the vehicle in front could act as context data. To make this multitude of data analyzable, data must be cleaned and fused in data pipelines. Cleaning happens in multiple stages, and requires decisions which have direct effects on patterns in the data. Time series data are often up- or down sampled, potentially altering characteristics of signals of interest. Sharing the data pipeline alongside an uncleaned version of the data therefore should be the default when publishing research results. Data science has developed a number of solutions to store and document data and data pipelines, whose benefits and costs will be discussed in this talk. These approaches can be structured in three interdependent dimensions: data storage, data processing, and competencies required by developers and users of data pipelines. Data from empirical studies can be very challenging to store, process, and document. Solutions to these issues do exist, but they require a training which is yet to be implemented in the typical Psychology curriculum.

Trotz der rasanten Fortschritte bei der Fahrzeugautomation, lässt der heutige Stand der Technik im Massenmarkt noch kein autonomes Fahren zu. Obwohl Teil- und Hochautomation zunehmend Aufgaben übernehmen, die vorher durch den Mensch ausgeführt wurden, wird er deshlab auf absehbare Zeit relevant für die Fahrzeugführung bleiben, zum Beispiel als Systemüberwacher oder als letzinstanzlicher Entscheider im Falle von Systemversagen. Die Bewertung von Verhalten, besonders in der Interaktion mit Technik, wird durch die komplexeren Systeme anspruchsvoller, und erfordert die Einführung neuer Methoden der Nutzermodellierung. Bisher hat sich die Verkehrspsychologie auf die Evaluation manuellen Fahrens konzentriert. Die hierfür entwickelten Metriken, zum Beispiel für Ablenkung oder Müdigkeit, sind für automatisiertes Fahren jedoch nur bedingt relevant. Ziel muss es sein, die Fahrer-Fahrzeug-Interaktion so zu modellieren, dass belastbare Aussagen über das Verhalten des Gesamtsystems getroffen werden können. Zu diesem Zweck wurde der Modellierungsansatz DriveGOMS entwickelt, mit dem sich beliebige reale oder hypothetische Szenarien der Mensch-Fahrzeug-Interaktion modellieren lassen. DriveGOMS kann als formale Sprache angesehen werden, die einen niedrigschwelligen Zugang zur Fahrermodellierung erlaubt und beruht auf dem dem Goals, Operators, Methods and Selection Rules-Framework (Card, Moran & Newell, 1983). Mit DriveGOMS wird ein Baukasten von Verhaltenseinheiten zur formalen Beschreibung von Fahrerverhalten bereitgestellt. Diese Verhaltenseinheiten wurden in empirischen Studien (Realfahrt- und Fahrsimulatorstudie) aus den Blick- und Verhaltensdaten realer Fahrer (N=12) extrahiert. Sie ermöglichen die Modellierung reiner Überwachungstätigkeiten, ebenso wie die Modellierung vollständiger Kontrollübergaben. Ergebnisse der Modellierung mit DriveGOMS können Sequenzdiagramme zur Abbildung des Fahrerverhaltens, Label für vorhandene Daten, oder Vorhersagen für mögliche oder notwendige Abfolgen von Fahreraktionen sein. Damit schafft das Framework DriveGOMS eine bisher nicht dagewesene Möglichkeit, Fahrerverhalten für ausgewählte Situationen in Bezug auf individuelle Fragestellungen zu modellieren. Um einen plastischen Eindruck der Möglichkeiten von DriveGOMS zu gewähren, wird im Vortrag eine Modellentwicklung an einem relevanten Szenario beispielhaft dargestellt.

With automated driving being on the cusp of wide spread market introduction (Rivera & van der Meulen, 2014), Human Factors aspect of design and evaluation of the exact behaviour of automated vehicles gains crucial importance. Safety margins determine an envelope of possible trajectories for the automated vehicle, but as of today, the parameterisation of the vehicle’s behaviour within that envelope to create attractive driving styles has has not received wide spread attention yet (Scherer et al., 2015). However, for acceptance of automated driving functions, deemed critical to deliver the promised reduction in accident numbers, it is essential to design automated driving styles that win end users over to activate them on public roads. In two consecutive studies we investigated how to measure preferences for automated driving styles, whether or not drivers prefer being driven similar to their own driving, or if at least default styles can be created that a majority of the users enjoy. In a first study 43 subjects drove in three scenarios on a two-lane motorway in a motion-based driving simulator. Users were instructed to drive at 120 km/h, forcing them to overtake slower vehicles on the right lane. The scenarios were varied regarding the presence and behaviour of faster cars on the left lane, compelling subjects to decide on timing and execution of overtaking manoeuvres. Based on the data from this study, four prototypical driving styles were extracted, using a multivariate time-series clustering algorithm (Griesche et al., 2014). In a second study, 35 subjects from study 1 rated the attractiveness of the prototypical driving styles gained from that study, in addition to their own style. Using a Best-Worst-Scaling technique, preferences for either one of the prototypical driving styles or their own style could be measured. The results indicate that i) many but not all subjects do like their own styles, but not in every situation, and ii) certain styles exist which are preferred by the great majority of the users.

Modelling human drivers as discrete states is a frequently used approach to understand and analyse driver behaviour, but a great challenge lies in linking empirical data with proposed states (Toledo, 2003). We propose a modelling approach based on a fine granular analysis of the driving task, using the established GOMS methodology (e.g. John & Kieras, 1996). At its core, GOMS assumes goals whose fulfilment are the end point of any human action. To reach these goals, and transfer current states into goal states, operators are employed. Being mere constructs, goals of drivers while operating their vehicle in highly complex environments cannot be measured directly, but must be modelled. One possibility are task models using hierarchical abstractions of the driving task (Walker, Stanton & Salmon, 2015). However, unlike tasks in highly controlled environments, driving seems to be comprised of a multitude of parallel goals, with drivers enjoying a high degree of freedom in setting these goals in specific situations. In a simulator experiment with 22 subjects, we therefore explored a second possibility: To ask drivers what their concrete goals were, employing the Thinking Aloud Technique (e.g. Nielsen, Clemmensen & Yssing). Driving in either a highly controlled traffic scenario, or in highly complex traffic situation on a two lane highway, subjects were instructed to report their current goals, current and planned actions, as well as general thoughts throughout the drive. Both video and audio from these trials were recorded, and played back to the drivers after each drive, allowing to ask the subjects further specific questions about their goals and actions in specific situations. Categorising the goals from the recorded audio and video data then allowed us to use them to construct task models, based on the DriveGOMS-methodology (Käthner, Andrée, Drewitz & Ihme, 2016). The goals served as the endpoints of windows in which driving operators were applied, enabling us to construct discrete states directly based on empirical data.

Driving behavior prediction at roundabouts is an important challenge to improve driving safety by supporting drivers with intelligent assistance systems. To predict the driving behavior efficiently steering wheel status was proven to have robust predictability based on a Support Vector Machine algorithm. Previous research has not considered potential effects of surrounding traffic on driving behavior, but that consideration can certainly improve the prediction results. Therefore, this study investigated how different surrounding cyclists impact driving behavior of an ego car. A simulator study was conducted to collect driving behavior data of ego car drivers in the scenarios with different surrounding cyclist position settings. The impact of the surrounding cyclists on the ego driver behavior was found: When there were surrounding cyclists, the recognition rate of ego driver behavior patterns reached 100% later than when there was no surrounding traffic. In conclusion, driving behavior pattern recognition at roundabouts is impacted by surrounding cyclists, and the impact can be expressed in a quantitative way.

Abschlussbericht für das AP5 (Empirical Human Factor Analysis Techniques and Tools) des EU-Projektes HoliDes.

Die Anpassung des Fahrstiles einer Automation an Fahrerpräferenzen gewinnt im Rahmen des hoch- und vollautomatisierten Fahrens immense Bedeutung. Wir stellen eine generische Methodik vor, wie solche Fahrstile und Präferenzen von Fahrern dafür empirisch ermittelt werden können.

As multitasking in the real world often does, driving a car manually requires an orchestrated effort of motoric, perceptual and cognitive faculties. As Neisser already pointed out with the Perception-Action-Cycle (1967), actions must be described and explained both regarding these faculties and the situative context they are rooted in. Especially highly dynamic environments such as the domain of public traffic require using and integrating manifold data sources. To this end, the Critical Path Model - Goals, Operators, Methods and Selection Rules (CPM-GOMS; John & Gray, 1995) provides a strong methodological and theoretical framework. It allows to understand, explain and predict driver actions regarding their time course as well as in their context. In an explorative study, one female and two male subjects (23-33 years of age) drove a passenger car on a two and a three lane motorway, without receiving explicit instructions. Based on video and vehicle data we modelled drivers’ motoric actions, perception and cognition for different lane change manoeuvres. For this we employed the CPM-GOMS-approach as analysis and modelling technique. Here, we present our approach and first models.

Moderne Assistenzsysteme stellen einen integralen Bestandteile heutiger und künftiger Fahrzeuge dar. Um Sicherheit und Komfort durch eine Unterstützung des Fahrers bei der Fahraufgabe zu erhöhen, ist es häufig notwendig, Fahrerverhalten adäquat vorherzusagen. Gerade in Kreisverkehren wäre eine solche Unterstützung durch Assistenz wünschenswert. Obwohl diese Verkehrsknoten durch eine geringere Unfallrate als lichtsignalgeregelte Kreuzungen auszeichnen, sind verletzliche Verkehrsteilnehmer wie Radfahrer und Fußgänger überproportional von Kollisionen mit Fahrzeugen betroffen. Das Fahrerverhalten trägt hierzu wesentlich bei. Die Erkennung und vorhersage von Fahrerverhalten ist daher eine wichtige Herausforderung, um intelligente Assistenz und Automation zu realisieren, die hier im Falle von Fahrer-fehlverhalten unterstützend eingreifen kann. Ziel unserer Arbeit ist somit, genau diese Voraussetzung zu gewährleisten und geeignete Methoden wie etwa statistische Modelle für Verhaltensvorhersage an Kreisverkehren zu entwickeln. Ein entsprechendes Modell muss in der Lage sein, frühzeitig einen Abbiegevorgang oder aber den Verbleib eines Fahrzeugs im Kreisverkehr vorherzusagen. Um eine Datengrundlage für ein solches Vorhersagemodell zu schaffen, wurden in einer Feldstudie an drei lokalen Kreisverkehren Fahrverhaltensdaten von sieben Probanden gesammelt. Dafür verwendeten wir ein speziell ausgestattetes Fahrzeug, mit dem Daten über Umgebungsverkehr, den Fahrer und das Fahrzeug aufgezeichnet werden konnten. Für die Bestimmung der Fahrzeugposition kam ein differentielles GPS zum Einsatz. Kopf und Blickdaten wurden mit Hilfe des kamerabasierten Blickregistrierungssystems Smart Eye Pro 6.0 erhoben. Verschiedene Variablen wurden hinsichtlich ihrer Eignung zur Verhaltensvorhersage geprüft. Da Fahrerverhalten gut als Mustererkennungsproblem modellierbar ist, wurde dafür eine Support Vector Machine genutzt. Dabei erwiesen sich Stellung und Geschwindigkeit der Bewegung des Lenkrads als sehr gute Indikatoren für einen Abbiegevorgang bzw. den Verbleib im Kreisverkehr. In einem Abstand von ca. 10 m vor der jeweiligen Ausfahrt aus dem Kreisverkehr konnten in über 90 % der Fälle korrekte Verhaltensvorhersagen getroffen werden. Unsere Ergebnisse zeigen somit zum einen die gute Eignung des gewählten Klassifikators Support Vector Machine als Vorhersagemethode. Zum anderen heben sie die Notwendigkeit einer genaue Untersuchung verschiedenster Fahrparameter hervor, um Fahrerverhalten adäquat verstehen und vorhersagen zu können.

Object properties that can be perceived simultaneously and without focused attention are called features (Treisman, 1980). Examples include color and the movement of objects. To date, the question whether velocity is a Feature or not has received little attention. In order to build exact cognitive models of visual object recognition this research gap has to be filled. Therefore, the current study compared the visual perceptibility of color and velocity changes. A varying number of moving and non-moving squares was presented to 22 participants. In conditions with moving objects squares moved on linear trajectories with different velocities. Further, the color of the squares was varied between conditions, with the squares being either unicolored or multicolored. Based on paradigms from working memory research, the first presentation of the stimuli was followed by a brief pause and a second presentation (Luck & Vogel, 1997). In the second presentation color, velocity or both attributes of one square were changed. The participants had to decide whether both presentations were identical or differed. From the number of correct and incorrect answers the sensitivity index

Driving is a complex task requiring processing on psychomotor, perceptual and cognitive level. An intimate understanding of the occurrence and timing of the processing steps is necessary when modelling driver behaviour and intent in cognitive architectures (e.g., ACT-R, Anderson et al., 2004). However, until now, little research on the decomposition of the driving task into the involved operations has been conducted. The approach of Critical Path Modelling Goals Operators Methods Selection rules (CPM-GOMS, John & Kieras, 1996) provides a framework for describing motor, perceptual and cognitive operators of a task on a granular time level. Thus, we employed CPM-GOMS in order to analyse lane change manoeuvres accomplished on a two-lane highway in real traffic. Here, we present the decomposition of three different types of lane changes (merging to the highway, change from right to left lane, and back) based on video observation of driver and environment as well as driving data analysis of three participants driving on the German autobahn A39. The resulting model forms the basis for understanding the aspects of vehicle operations and serves the creation of cognitive models.

This paper focuses on modelling driver intention and behaviour at roundabouts in order to provide information on whether or not a driver intends to leave the roundabout when approaching an exit of a roundabout. Advanced Driver Assistance Systems’ effective work depends on adequate driving intention recognition and behaviour prediction, so if the driver intention and behaviour at roundabouts can be modelled and predicted, the roundabout safety and efficiency can be significantly improved. As the driver intention recognition is basically a pattern recognition problem, the machine learning theory is a good candidate for training models. The data for training and validating the models has been planned to be gathered from both the laboratory controlled simulation and field study.

Using data from a car following field study, we present a preliminary evaluation of the Driving-by-Visual-Angle car following model (Anderson & Sauer, 2007). The results show that the DVA-model offers no improvements over existing car following models.

The paper introduces the method and tool CONFORM (Conflict regocnition by image processing methods). CONFORM will be integrated as a driver model into the ibeo test vehicle during the project phase of EU–project HoliDes. The aim of CONFORM is to support the system designer to properly parameterize the default behavior of a highly automated vehicle to guarantee a high system acceptance. Thereby CONFORM addresses intra and inter individual differences in the driving behavior. CONFORM measures the difference between the default system behavior and the natural driving behavior of a human driver situation-dependent to determine the necessity of an adaptation. Based on a driving simulator study the paper describes how CONFORM is able to visualize and to cluster certain driving patterns/styles in a vehicle following/vehicle approaching scenario. We use the study results to derive recommendations for the design of the system behavior of highly automated vehicles.

Durch moderne Assistenz- und Informationssysteme (ADAS) verändert sich die Fahraufgabe stark – wie genau ist weitgehend unklar. Eine empirische Einschätzung des Einflusses dieser Systeme ist meist nicht zielführend, weil die dafür eigentlich nötigen Ressourcen nicht zur Verfügung stehen. Mit ausführbaren Modellen des menschlichen Fahrers ließe sich die Wirkung etwa eines Spurwechselassistenten wesentlich besser untersuchen und verstehen. Ein solches Modell existiert bisher nicht. Die Grundlagenforschung der Psychologie stellt kognitive Architekturen bereit, die Theorien über die aufgabenunabhängigen Bestandteile menschlicher Kognition als Computerprogramme implementieren. Im Rahmen einer solchen Architektur sollen ausgewählte Bereiche des Fahrerverhaltens modelliert werden.