4.2.1. Situational awareness

The situational awareness (SA) of an individual, as defined by Endsley [END 95], is made up of three levels:

• – level 1: perception of the elements in the environment. At this level, the information is not interpreted. Information is simply received in a raw, unprocessed format. This level contains information about the states, attributes and the dynamic of the elements that are relevant to the environment;
• – level 2: understanding of the current situation. This level intervenes after perception, as soon as the data can be integrated into existing frameworks. Information processing, which implies understanding, consists of putting the characteristics of the ongoing situation into correspondence with frameworks in the memory, which represent prototype situations. This stage of understanding is important in apprehending the significance of the elements seen at the previous level, by means of an organized image of the situation, in relation to the objectives;
• – level 3: projection of future states. This level is associated with the ability to anticipate, meaning to project the states of the elements perceived and understood, into the near future. The exactitude of the prediction is strongly correlated to the exactitude of the two preceding levels. This is the level for forecasting the future state of the perceived situation.

In a general manner, research literature mentions two main states of SA:

• – it can be insufficient: in this case, human agents do not manage to construct an exact and complete representation of their environment, which is necessary to achieve the objectives that are assigned to them. This can therefore have consequences on the performances obtained;
• – it can be sufficient: SA is described as sufficient when a human agent perceives and correctly interprets information in the environment in which they find themselves. Thus, they can react in accordance with any unforeseen event, to restrict or cancel all the damage that this event may have brought about. This is the optimal state.

Sufficient SA of the human agent is one of the major challenges that “design for SA” (termed design for situational awareness by Endsley [END 16]), wishes to tackle. This is also the challenge that we wish to take on in our work. In addition to the SA, another selected objective is transparency of the controller.

4.2.2. Transparency

The Larousse dictionary proposes five definitions of the word “transparent”:

• – definition 1: describes a body that transmits light by refraction and through which the objects can be seen with clarity;
• – definition 2: describes a material that allows light to pass through it;
• – definition 3: describes a cloth, paper or skin that is thin enough so that we can see through it;
• – definition 4: with reasons or direction that are easy to guess;
• – definition 5: with a clear functionality that we do not seek to hide from opinion.

These definitions allow us to identify two broad approaches. In approach 1, definitions 1, 2, 3 indicate that an object or system is transparent if it is possible to see through it. In approach 2, relative to definitions 4 and 5, an object or system is transparent if we easily understand its operation. Even though these two approaches appear to be similar, they are functionally very different [OSO 14]. While approach 1 manages the aspect of form, approach 2 manages the aspect of foundation and we have given preference to the latter in our research.

Following this approach, transparent systems communicate information that allows how they function and/or what they do to be understood. In research literature about the interaction between a human agent and an automated system, several authors define the transparency of a system according to this characteristic. For example, Cring and Lenfestey [CRI 09] have suggested that transparency refers to the perceived predictability and the understanding that the human agent has of a specific behavior of automation. Kim and Hinds [KIM 06], for their part, specify that if the human agents do not understand the system’s operational logic, normal and even routine behaviors of the latter can then be perceived as errors by human agents. However, even though a high level of transparency (revealing several pieces of information about operation of the controller) is in general preferred¹ by human agents [SAN 14], too great a quantity of information can lead to an information overload and thus affect levels 1 and 2 of the SA.

Some authors have worked on transparency of the controller in the context of driving automobiles. For example, to ensure a successful takeover, Eriksson and Stanton [ERI 15] have proposed that Grice’s [GRI 75] four maxims should be used to increase the transparency of autonomous cars. The following feature among these maxims:

• – the quantity maxim which stipulates that no unnecessary information must be added to the interface;
• – the quality maxim which stipulates that all the information sent by the automated system must be true;
• – the relational maxim which stipulates that all the information must be contextualized and pertinent with respect to the task that the controller is in the process of carrying out;
• – the manner maxim which stipulates that ambiguity must be avoided and that the information must be presented in a structured and brief manner.

Eriksson and Stanton [ERI 15] thus suggest that by using these maxims in the interface design of a highly automatized car, a good mental representation of the controller’s operation will be produced in the human agent, and thus the transition phase between the automatic mode and the manual mode will be facilitated. Naujoks, Forster, Wiedmann and Neukum [NAU 16] looked at the transparency of highly automatized cars involving communication about future maneuvers. Thanks to their results it has been determined that the vocal method made it easier to extract information that is relevant for understanding the controller’s intentions.

While not much of this research has been carried out, there are models of transparency that exist and which are proposed by certain authors, which can be applied to automobile driving. Among these models there is, for example, the model of Situation Awareness-based Agent Transparency and Lyons’ models, which are the one that we have used (see Figure 4.1). Lyons looked at the question of transparency from two angles: the transparency of a robot for humans and the transparency of humans for robots; where here the robot refers back to an automated system. The transparency of the robot makes reference to the set of information that the robot must communicate to the human [LYO 13].

This transparency is based on various models: the intention model, the task model, the analytical model and the environmental model.

• – The intention model: here the intention refers to the general objective for which the robot was designed, or even to its overall objective. At this stage, two questions arise: what is the robot’s “raison d’être”? Macroscopically, what functions is it able to carry out? To distinguish “the intention in action” from “the intention in a general sense” defined during the design phase, it appeared more judicious to rename this model “the overall objective model” (of the robot). In the following, in order to avoid any ambiguity with an intended maneuver, we will refer to a “model of the general objective”.
• – The task model: Lyons specifies that the human agent analyzes the robot’s actions in line with a specific cognitive schema². The task model therefore aims to provide details to help the human agent establish this cognitive schema. In particular, it must contain information that makes it easier to understand a given task, information regarding the robot’s objective at a given instant in time, regarding the robot’s advancement with respect to its objective, its advancement relative to the tasks that can be carried out by the robot and knowledge of the errors that could occur while the task is being carried out. These various pieces of information allow a shared representation to be established, between the robot and a human agent, of actions that must be accomplished for a given task. Communication of the robot’s intention in relation to the objective that it wants to achieve allows the human agent to find out the state of advancement of the task as well as the reason why it carries out a given action or adopts a given behavior. Lyons suggests that communicating this information improves the human agent’s representation of the robot and also helps to improve surveillance of the latter.
• – The analytical model: Lyons specifies that to achieve the objective that is assigned to it during its design, a robot must acquire and analyze a large quantity of data. Given the complexity of this information, the human agent may have difficulties understanding how the robot makes its decisions. The analytical model therefore aims to communicate to the human agent the analytical principles used by the robot during decision-making. This is very useful, in particular in complex situations in which the level of uncertainty is high.
• – The environment model: in difficult and potentially hostile conditions in which time constraints are high (such as military situations), it is essential that robotic systems operate with a dynamic that is in sync with their environment. The robot must communicate to the human agent the understanding that it has of topographical variations, meteorological conditions, threats and time constraints in a particular environment. This type of information indicates to the human agent that which the robot exactly perceives and this will improve the SA that the human agent has of its environment. Moreover, knowing that the robot knows the environmental conditions will help the human agent calibrate the confidence that can be attributed to it.

In addition to the transparency of the robot for the human agent, the transparency of the human agent for the robot must be taken into account. This transparency refers to the set of information that the robot can integrate concerning the state of the human agent, as well as the information shared between the two decision-makers to manage and understand the sharing of work [LYO 13]. This transparency is therefore based on two models.

• – The teamwork model: the robot and the human agent form a team in which each must be able to clearly identify the objectives that are assigned to them. In this model, which we have named the “cooperation model”, the automation must indicate to the human agents the tasks that they are responsible for, the tasks for which the human agent is responsible and at what level of autonomy they operate. This information will allow the human agent to predict the actions of the automation.
• – The human agent model: this is the model that will allow the state of the human agent to be analyzed. Thanks to this model, the robot may be capable of diagnosing the level of stress, cognitive overload, etc. of the human agent. If the robot identifies cognitive overload in the human agent, Lyons suggests that it responds with an increase in the degree of automation for as long as this state is observed and particularly when it is carrying out critical tasks.

Although structuring in sub-models of Lyons’ model is noticeable, the author does not specify, among all this information, which the most important are. Indeed, to cooperate, the agents must understand and share a set of information that constitutes the common frame of reference [DEB 09]. This information comes from exchanges at the “information gathering”, “information analysis”, “decision-making” and “action implementation” levels [PAR 00]. Moreover, Lyons’ models do not specify in which situation such or such information must be presented. Is this in all situations or in abnormal situations? In other words, they do not manage the question of information prioritization, considering the visual load or the context for example. Taking into account the four functions of Parasuraman and his colleagues [PAR 00] seems important in order to answer these questions and to complete the structuring proposed by Lyons’ models.

4.3.1. Presentation of the approach

Since our research has been carried out with a view to cooperation between human agents and autonomous vehicles, we have paid particular attention to establishing a common frame of reference between the two agents. This is the reason why we have transposed Lyons’ models [LYO 13] into the driving domain, using the function proposed by Parasuraman et al. [PAR 00]. This has enabled several principles of transparency to be distinguished in order to connect them with each of these functions. In order to obtain solid information to display, we identified the information requirements of the human agent in manual mode thanks to the first two stages of cognitive work analysis. These were then structured according to Michon’s three levels of control [MIC 85]: a strategic level, a tactical level and an operational level, taking the various display possibilities into account. Once the question of information content was dealt with, creative sessions took place to direct the way this could be presented in augmented reality. In order to direct the display of information as a function of the various available displays and the driving context, rules of display – also known as “rules of supervision” – were defined for the human–computer interactions (HCI) supervisor. Lastly, an experimental validation was designed in order to validate certain principles, since the time available was too restricted in order for them all to be validated. In the following section, we will present the way in which the principles of transparency have been defined.

4.3.2. Definition of the principles of transparency

The school of thought that is advocating human–computer cooperation considers that the human agent and the technical agent must operate in a team, meaning by cooperating to have the best possible performance, while facilitating the work of the other. A common frame of reference must therefore be maintained and/or constructed between these agents, where this frame of reference can be supported by an HCI, which corresponds to what is known as common work space. This CWS can be composed of several attributes such as [DEB 06]:

• – the formulation of information arising from activities of information acquisition;
• – the formulation of problems that arise from diagnosis activities;
• – the formulation of strategies that arise from schematic decision-making activities;
• – the formulation of solutions that arise from precise decision-making activities;
• – the formulation of instructions that arise from solution implementation activities.

In Lyons’ models, there is no clearly explained correlation between each sub-model and the four functions of information processing in Parasuraman et al. [PAR 00] or the attributes of the CWS, about which we have just given a reminder. Thus, in our work, direct connections between the principles of transparency arising from Lyons’ models, and these functions, will be established to supply this space.

Moreover, to ensure complete consistency of the approach and to take into account our research topic of autonomous driving, when it becomes possible we will establish relationships between each principle and Michon’s model [MIC 85] which defines automobile driving using three levels: strategic, tactical and operational. In order to have an analysis matrix, we have therefore paired the functions of Parasuraman et al. [PAR 00] and Michon’s levels [MIC 85].

Table 4.1 presents this matrix which we have named the IPLC-based matrix, meaning Information Processing and Level of Control-based matrix, that is, a matrix based on information processing and the levels of control.

In Table 4.1, the letters S, T and O refer respectively to the strategic, tactical and operational levels. Concerning the abbreviations I.T, I.A, D.M and A.I, they refer respectively to the functions “information gathering”, “information analysis”, “decision-making” and “action implementation”. We have defined 12 principles on the basis of Lyons’ models, concerning the LAR project in which the main functions of the controller have been to carry out a change of lane, to adjust its speed and its distance as a function of the speed and the distance between the autonomous vehicle and the surrounding vehicles. For each of them, we have established, where possible, the connection with the IPLC matrix by putting 1 in the corresponding box.

These principles are represented in Figure 4.2.

While the principles that we are going to describe are general and are correlated with Michon’s three levels of control; no particular attention was paid at the strategic level because it is outside the field of investigation of the LAR project. For each of Lyons’ sub-models, we are going to give an example of a defined principle.

4.3.3. Cognitive work analysis

Cognitive work analysis (CWA) is a methodology that is typically used to design interfaces described as ecological. These interfaces have the objective of making the constraints and the possibilities of actions visible in a given field of work and facilitating the use of a low-cost cognitive control method based on automations (skill-based) or on rules (rule-based) rather than on knowledge (knowledge-based). Their primary objective is to help the operators face new situations in complex socio-technical systems [VIC 02]. Their development relies on three stages of the CWA: work domain analysis, task analysis and competencies analysis.

The ecological interfaces were initially designed to support control activities for processes (nuclear, petrochemical) or driving. For a long time, they have also been developed to facilitate the supervision of autonomous systems: inhabited submarine vehicles [KIL 14] as well as autonomous vehicles [REV 18]. In the context of the supervision of an autonomous vehicle, these interfaces have different aims depending on the phase to which they are addressed: takeover phase or autonomous driving phase. In the first case (documented by Revell et al. [REV 18]), they are designed to make it easier to manage an unexpected situation, whereas in the second case (documented by Kilgore and Voshell [KIL 14]), they have the primary objective of making it easier to understand the operation of the system and thus join with the objective of transparency. Our work comes closer to the study by Kilgore and Voshell, since it relates to the autonomous driving phase. The takeover phase, to be carried out correctly, requires the human agent to have a correct mental representation of the environment and of the behavior of the autonomous agent.

In the following sections, we present the results from two analyses that have been carried out: analysis of the field of work and analysis of the control task.

4.3.3.1. Work domain analysis

Work domain analysis was proposed by Rasmussen [RAS 86]. Due to the formative nature of cognitive work analysis, work domain analysis focuses on the constraints related to the safety and performance of a complex system rather than considering specific scenarios [NAI 01, NAI 13]. These constraints can be represented by an abstraction hierarchy (AH) that includes five levels which are, in decreasing order of abstraction: domain purpose, values and measures of priority, functions associated with the purpose (or general functions), processes associated with objects (or physical functions) and physical objects. The links between the levels in the abstraction hierarchy express “means–ends” relationships, also known as “how–why” relationships. Effectively, the connections between a target function and the lower levels of abstraction indicate how a function is operationalized. Inversely, the links between a target function and the higher levels of abstraction indicate why this function exists.

4.3.3.1.1. Abstraction hierarchy

An abstraction hierarchy has already been implemented in the field of car driving. Several research publications have looked at this topic (see, for example, [SAL 07]). Concerning our issue, in Figure 4.3, we present an extract of the analyses made using many literature articles such as [JEN 08], interviews with LAR project collaborators and several iterations of which the first are consultable in [POK 15a, POK 15b].

In Figure 4.3, the information is distributed over several levels:

• – domain purpose: the system’s “raison d’être”. In the case of the task of driving, the identified purposes are: road transport that is effective, comfortable and completely safe;
• – values and priority measures: these correspond to the principles, priorities and values to follow in order to achieve the domain purposes. In our case, it is a case of complying with the defined itinerary, the separation distances between vehicles, speed limits and avoiding collisions between vehicles;
• – functions related to purposes: also known as general functions, they represent all the main functions that the system must carry out. We have identified five functions, including the function “monitoring vehicle maneuvers in the driving environment”, for example;
• –functions related to objects: these functions, also known as physical functions, correspond to functions supporting general functions. For example, the function “detect and recognize speed panels” has been highlighted;
• – physical objects: these are sensors, mechanical parts and/or algorithms that guarantee the functions that relate to the objects. For example, navigation systems.

4.3.3.1.2. The definition of the information requirements for driving using work domain analysis

We have modeled the abstraction hierarchy (AH) for the driving domain independently from the agent that carries it out, whether a technical or human agent. As Li and Burns [LI 17] have specified, a function in the AH can be the responsibility of the technical agent, the human agent or both (shared control). In other words, the AH can be a basis for the allocation of functions. In their research, Li and Burns have used the AH in the field of financial transactions in order to highlight two scenarios of function allocation, one that represents low automation and the other high automation [LI 17]. These authors have also specified that pointing out the function relating to objects (physical functions) on an interface allows the losses of situational awareness in the human agent to be restricted.

This result is of particular interest, because it allows us to identify the potential information that will then lead to the definition of the information to be presented effectively in order to maintain/establish sufficient situational awareness of the human agent. This identification is carried out from physical functions determined in the abstraction hierarchy. The speed panels and the action of changing lanes constitute examples of potential information.

Work domain analysis has allowed us to define the general information requirements of the driving task, without concentrating on particular situations. However, given that our research attributes particular attention to the maneuver of changing lanes, it is necessary to identify information requirements relating to this maneuver in a specific manner.

Therefore, we have carried out the analysis of this maneuver by using the control task analysis (ConTA), in order to identify the information-processing activities performed by the human agent. We present this analysis in the following section.

4.4.1. Interfaces

The objective of our work is to validate or invalidate the principles of transparency that are defined in the context of the use of an autonomous car in level 4 of the SAE taxonomy and, consequently, to evaluate the optimal level of transparency for an HUD interface that is specific to the LAR project. Therefore, we have decided to activate or deactivate the display of information associated with these principles following the four information processing functions in Parasuraman et al. [PAR 00]: “information gathering”, “information analysis”, “decision-making” and “action implementation”. A total of 16 interfaces can thus be defined. However, we have only focused on five interfaces among the 16 for two main reasons:

• – time constraints: since human resources related to the LAR project are limited and time is limited, it was not possible to develop each of these 16 HCIs;
• – financial constraints: having 16 HCIs to evaluate requires a large number of participants to be available in order to validate them.

Table 4.4 sets out the functions present in each of the HCIs that have been selected and developed.

In emphasizes basic information; Iw emphasizes information related to information gathering and action implementation; Ii is like Iw with the additional information related to situation analysis; It emphasizes information related to information gathering, information analysis and decision-making; Is presents all information.

In does not contain any information from the principles of transparency in relation to the functions defined by Parasuraman et al. [PAR 00], as shown in Table 4.4. However, it is not lacking in information. Indeed, we have identified the basic information that makes up this interface and that makes it the reference interface. Identified on the basis of the analyses of interfaces of recent vehicles in circulation and certain prototypes of autonomous cars, this information includes in particular:

• – the instantaneous speed of the autonomous vehicle;
• – the speed limit defined for the section in which the autonomous car is located;
• – an indication that the autonomous mode is active;
• – the duration of the autonomous mode before entering a zone that requires a takeover;
• – navigation: this is not as dynamic as it is with a GPS. It provides an insight in the form of a map of the autonomous car’s journey, with an indication of the local journey to be taken in the event of a motorway split or exit from the motorway;
• – the blind spot: this indicates that another vehicle is present outside of the autonomous car’s range of vision.

Concerning the HCI evaluations, they will be based on the differences observed during data analysis. Thus:

• – the difference between Iw and Ii will allow the added value of information from the “information analysis” function to be evaluated;
• – the added value of information from the “decision-making” function to be evaluated;
• – the difference between Is and It will allow the added value of information from the “action implementation” function to be evaluated.

The various symbols and annotations used for these interfaces have been defined after several creative sessions. Some of them are presented in [POK 16].

4.4.2. Hypotheses

During manual driving, the human agents construct a representation of the situation (this is situational awareness) on the basis of elements relating to their perception, their understanding and the projection that they make of it. This representation must be relevant, meaning that it must allow them to acquire the right information at the right moment in time. In the context of the autonomous car, the HCI that we have defined all aim to facilitate perception but in addition, for some of them, understanding of the situation and projection of its state into a near future, such as the fact, for example, that the autonomous car is going to take an exit. Therefore, we make the hypothesis that the level of transparency of the HCI’s is going to have an impact on the SA of the drivers (hypothesis 1). Moreover, we also make the hypothesis that the least transparent HCI (HCI 1) will not be the most appreciated of the human agents and that the latter will have a preference for interfaces that clearly show the actions that the controller carried out or is going to carry out (hypothesis 2).

4.4.3. Participants

A total of 45 people, ranked as per the distribution in Table 4.5, have taken part in the experiment that took place in the DrSIHMI³ simulator. The average age of the 23 women was 43 years with a standard deviation of 9.23; the age for the men was 43.5 years with a standard deviation of 9.53. The duration in years for which they have held a license was 22.4 with a standard deviation of 10.33.

For each participant, the test lasted for approximately two hours, with effectively 40 minutes of driving time. In a few cases, the two hours were exceeded, in particular when the participant spent more time in debriefing interviews carried out after the tests. In any case, we set the maximum duration of the test at two hours and thirty minutes at the time when the study started for each participant in order to avoid any tiredness. This limit was complied with.

4.4.4. Equipment

The study was carried out in the simulator at the Technology Research Institute System X (IRT SystemX), the DrSIHMI (Figure 4.5). This static simulator is made up of the following elements:

• – a car driving seat: this includes a steering wheel, the accelerator, brake and clutch pedals, a gear stick, a button box, a non-AR HUD, an AR-HUD and a remote display:
  • - the button box allows the driver to show the degree of discomfort felt during a situation, with buttons 1 and 2,
  • - the non-AR HUD: given that DrSIHMI does not have a physical dashboard, this display, with dimensions 15*4°, allows classic information to be displayed such as the speed of the ego-car,
  • - the AR HUD: with dimensions 15*5°, this display allows virtual information to be projected in the driving environment for the perception of augmented-reality elements,
  • - the remote screen contains information about the strategic level. This screen presents an image of the journey for the autonomous vehicle to follow, as well as the distance to the arrival point;
• – a curved projection screen with an opening of 180°.

In this equipment context, and in order to carry out the experiments, several scenarios have been developed.

4.4.5. Driving scenarios

A driving scenario sets out a specification of all the events that can occur in the environment. These are, in particular, traffic vehicle maneuvers, modification of the road topology (e.g. bend or straight road) or weather conditions. In the context of our work, the scenarios have been developed by specific technical resources on the basis of specifications that we have provided, thanks to the software SCANeR. Developed by OKTAL SA, SCANeR models scenarios use graphics involving “condition-action” pairs (“if…, then”) in the driving environment [REY 00]. The driving scenarios are designed as a succession of driving scenes in a given environment or terrain.

Figure 4.6 shows a bird’s eye view of the terrain used.

We have defined several scenes: normal scenes such as the change of lane of the autonomous vehicle due to a separation of motorway, and less everyday scenes such as overtaking two heavy goods vehicles combined with the arrival of a police vehicle.

Before the beginning of each test, it was specified to each participant that the behavior of the autonomous car may not be homogeneous for the duration of the scenario. Effectively, we have specified that the autonomous vehicle may accelerate and brake abruptly, and that it may in certain circumstances reduce the security distances. This presentation phase, carried out on a computer, mainly aimed to comply with the principles arising from the model of the general objective⁴. These principles aim to make the human agent understand what the autonomous car is and what it can do. Varying the relative order of the scenes has allowed four different scenarios to be defined:

• – scenario 0: this lasted for three minutes and thirty seconds and constituted the learning scenario;
• – scenarios a, b and c: of an average duration of nine minutes, these similar scenarios (driving scenes that are identical but presented in a different order) made up the scenarios for each drive. They were associated with one of the five HCIs and submitted in a different order to each participant.

The numbers of participants for each interface and for each scenario are presented in Table 4.6.

During each of these scenarios, various measures have been carried out.

4.4.6. Measured variables

In the context of these experiments, we have collected data about the cognitive activities of the participants (information gathering and situational awareness) on the one hand, and about elements relating to the user experience (satisfaction) on the other hand. Table 4.7 presents all the variables studied as well as their possible values (for discrete quantitative variables) or their means (for qualitative variables).

The dependent variables have been collected at different times:

• – answers to questionnaire Q1 regarding the situational awareness were collected only in the first driving number (DN 1), on one of the following two scenes: the first corresponds to an overtaking of a train of vehicles with a lane change of one of them; the other corresponds to an overtaking of two heavy goods vehicles with the arrival of a police vehicle. They aim to evaluate the situational awareness;
• – classification of the HCI into first, second and third positions was carried out at the end of the third experimental condition. This classification aims to evaluate the acceptability of the HCI.

In the experiment validation, we have manipulated four independent variables or “factors”: the driving number (DN), the interface (I), the freeze situation (F) and the driving scenario (S). Moreover, there were control factors (or independent variables) from questionnaire Q0 about driving habits such as the age of participants (A) and their gender (G).

4.4.7. Statistical approach

Given that several variables have been measured, these have all been considered through the bias of a multivariate approach, instead of carrying out univariate analyses several times. In the family of multivariate analysis tools, the classifications and regrouping techniques are well-known [JOH 07]. In the context of our research work, we have opted for the classifications because they present the advantage of “easily” showing the relations between the variables and the effects of each of the variables [BEN 92]. The presence of quantitative variables and qualitative variables has led us to use multiple correspondence analysis (MCA) in place of the principle component analysis (PCA) that is generally used [BEN 92].

If the MCA is used more in the analysis of qualitative variables, in particular in multiple choice questionnaires or enquiry data [BEN 92, CES 14, BIL 16], it is important to note that it leads, with quantitative variables coded in a fuzzy way, to a demonstration of nonlinear relations. Consequently, the models used in it are more complex than those of the PCA [BEN 92, LOS 14]. For a first analysis, the fuzzification model (FM) is used for arithmetical means and value intervals. The FM is less sensitive to extreme values (values that are sometimes erroneous) because the adjusted average is less sensitive to the abnormal values than the arithmetical average. Before going further, we will spend time on the following considerations:

• – consideration 1: to evaluate the impact factors, a summarizing operation is necessary. This summary requirement leads to a lower loss of information with local fenestration than without [LOS 01, SCH 15];
• – consideration 2: there are other models of scale change but this offers good results (see [LOS 14] for comparative studies);
• – consideration 3: each histogram and its corresponding window were carefully verified.

Applications of the MCA in the field of transport exist [CHA 13, CES 14, BIL 16] with qualitative variables [SCH 15] and with fuzzified quantitative variables. The independent variables that we have selected are as follows: the participant (P), the driving number (DN), the interface (I), the driving scenario (S) and the freeze situation (F). The data collected during each experimental condition that we present in this chapter is issued for each participant: from questionnaire G on the SA (G1–G6), from the ranking of HCIs in first, second and third positions (respectively FI, SI and TI). There were only 45 datasets for the measures of situational awareness (G) because these were only collected during the first driving number (DN1). Concerning the FI (HCI ranked in first position), SI (HCI ranked in second position) and TI (HCI ranked in third position), they were only provided at the end of the third driving number (DN3). In the following sections, we present the results obtained from this experiment.

4.5.1. Situational awareness

The situational awareness was evaluated during the first drive. During this drive, questionnaire G was administered when a freeze occurred (see Table 4.8). The questions asked were related to six items: vehicle action; vehicle dynamic; vehicle speed; future vehicle action; number of vehicles in front of the autonomous vehicle; and number of vehicles behind the autonomous vehicle.

The response to each of these questions (G1–G6) has been coded using two modes: 0 and 1. They correspond respectively to a false answer and a correct answer.

4.5.1.2. Discussion

We have observed a significant effect of variable I on G2 (dynamic of the vehicle). Regarding this question, the number of correct answers was very low in In and the highest number of correct answers was observed in Iw. In other words, the HCI that presents elements relating to the collection of information and the implementation of action allows the participants to have a better perception of the dynamic of the autonomous vehicle than one that does not present any of these elements. One of the specific elements in Iw that highlights the dynamic of the autonomous vehicle is the table of nine boxes, which tends towards validation of the principle of T1 transparency and which stipulates in part that “the driver must be capable of detecting actions (change of lane, continuation in lane, change of speed) that the car is in the process of carrying out, and understanding them”.

Similarly, there is a significant effect of I on G6 (number of vehicles behind the autonomous vehicle). Concerning this question, the number of correct answers was very low in Is and the highest number of correct answers was observed in Iw. In other words, with the HCI that displays the collection of information and the implementation of the action, the participants perceived the number of vehicles behind the autonomous vehicle better than with the HCI, in which not only the results of information collection and information analysis, but also the results of decision-making are presented. This result is even more surprising than in Is – there were explanations about the presence of a police vehicle or a vehicle arriving quickly from behind. We establish the hypothesis that this HCI, which presents numerous pieces of information, caused an information overload. Effectively, as Bawden and Robinson have mentioned [BAW 09], a large quantity of pieces of information about an interface can cause a reader to not read them all.

These two results tend towards the validation of hypothesis 1 according to which the transparency of HCIs causes differences in the representation of the situation that the agents set up. In summary, from the point of view of cognitive activities, Iw (HCI that shows the information relating to information collection and action implementation) leads to better results from the point of view of the situational awareness.

4.5.2. Satisfaction of the participants

The participants carried out three drives and therefore saw three HCIs. At the end of the three drives, the following question was asked of them, by way of conclusion: “by order of preference, which are the HCIs that have best allowed you to understand the system? In first position, second position?”. For example, a participant X who has seen In, Iw and Ii could choose to rank Ii in the first position, Iw in the second position and In in the third position.

4.5.2.1. Results

The data are different to the previous for two reasons:

• – reason 1: the values correspond to relative evaluations, whereas the previous values correspond to absolute values;
• – reason 2: the 45 participants have not all evaluated the same interfaces, with the exception of the In interface.

Therefore, a specific procedure has had to be used. The six graphs in Figure 4.8 clearly show that the In interface is appreciated least of all by the participants, in comparison to the four others.

For a more general comparison, Figure 4.9 shows – for each interface – the ratio between the number of times that it has been ranked in first position and the total number of times that it has been tested. For example, for the In interface, these two numbers are worth 9 and 45 respectively, which provides a ratio of 8.9 %.

Figure 4.9 shows that the Is interface is slightly better appreciated than the It interface.