An Empirical Investigation on the Visual Imagery of Augmented Reality User Interfaces for Smart Electric Vehicles Based on Kansei Engineering and FAHP-GRA

Abstract

Smart electric vehicles (SEVs) hold significant potential in alleviating the energy crisis and environmental pollution. The augmented reality (AR) dashboard, a key feature of SEVs, is attracting considerable attention due to its ability to enhance driving safety and user experience through real-time, intuitive driving information. This study innovatively integrates Kansei engineering, factor analysis, fuzzy systems theory, analytic hierarchy process, grey relational analysis, and factorial experimentation to evaluate AR dashboards’ visual imagery and subjective preferences. The findings reveal that designs featuring blue planar and unconventional-shaped dials exhibit the best performance in terms of visual imagery. Subsequent factorial experiments confirmed these results, showing that drivers most favor blue-dominant designs. Furthermore, in unconventional-shaped dial designs, the visual effect of vertical 3D is more popular with drivers than horizontal 3D, while the opposite is true in round dials. This study provides a scientific evaluation method for assessing the emotional experience of AR dashboard interfaces. Additionally, these findings will help reduce the subjectivity in interface design and enhance the overall competitiveness of SEV vehicles.

1. Introduction

By 2030, the number of cars worldwide is expected to increase from 1.3 billion to 2 billion [1]. The huge increase in the number of cars will bring enormous environmental pressure to the region and the world, especially in terms of air pollution and the greenhouse effect [2,3]. Recently, smart electric vehicles (SEVs) have undoubtedly become a hot research topic for the Sustainable Development Goals [4,5], which have great potential to facilitate the energy crisis and environmental pollution and appeal to many consumers by emphasizing the user experience. The annual sales volume of new energy passenger vehicles in China in 2023 has reached 7.736 million units, of which SEVs accounted for 6.619 million units, with a penetration rate of 85.6% [6]. As a way to reduce carbon emissions and improve driver experience, SEVs have a bright future in China and worldwide. With the development of augmented reality (AR) and In-Vehicle Information Systems (IVISs), in-vehicle AR display technology has been widely applied and researched as an important functional system for SEVs.

AR is an advanced form of Human-Computer Interaction (HCI) that provides intuitive and rich interface information by embedding and overlaying virtual elements onto the real environment [7]. The applications of AR technology are extensive, spanning various domains such as gaming, education, entertainment, and manufacturing [8]. In the automotive sector, AR applications are primarily categorized into two types: one is in-vehicle display systems designed to provide information to drivers, and the other is auxiliary systems used during the automotive development process [7]. This study focuses on AR dashboard display systems intended for drivers. An AR dashboard combines augmented reality technology with dashboard displays, overlaying driving data, navigation, and assisted driving information onto live road video and presenting it to the driver through the dashboard display [9]. If AR is appropriately integrated with real-time road video, it can enhance drivers’ situational awareness and thereby improve driving safety [10]. Some automakers have already adopted this technology in their production models. For example, China’s SAIC Group launched the Rongwei MARVEL X in 2018, which pioneered the rendering of visual recognition results, fused positioning results, and map navigation information into AR images displayed in the in-vehicle dashboard. In addition, many scholars have conducted research on in-vehicle AR for the human-machine interface (HMI). Calvi and D’Amico [11] found that in-vehicle AR warnings significantly enhance the safety of left turns. Liu and Yin found through eye-tracking experiments that the reading performance on blue AR interfaces was the poorest, while green and adaptive colors demonstrated the most stable performance [12]. Zhong and Cheng [13] studied how environmental illuminance, interface color, and speed font design affect driver visual fatigue and visibility. Li and Wang [14] examined the impact of AR interface color combinations on the visual search performance and cognitive efficiency of drivers, considering gender and driving scenarios. However, most HMI research on in-vehicle AR has focused on driver safety [7,11,14,15,16], neglecting the experiential and emotional aspects of the driver. User experience is a pivotal determinant in enhancing user engagement and overall usage [17]. It encompasses both the behavior and emotions that users exhibit towards a particular object or system [18]. In addition, the user’s emotional experience significantly affects purchase intention [19], usage intention, and user satisfaction [20]. Given its importance, this study aims to evaluate the emotional experience elicited by the user interface of an in-vehicle AR dashboard. Simultaneously, this study developed a method for evaluating the visual imagery and subjective preferences related to in-vehicle HMIs.

2. Theoretical Background

2.1. Kansei Engineering

Kansei engineering is a product design and development method based on human emotions and needs. Its core lies in the quantitative analysis of user emotions and feelings [21]. Kansei engineering focuses on capturing users’ “Kansei” during the design process, which refers to their perceptions of aspects such as the color and shape of products or interfaces. This approach emphasizes addressing users’ emotional needs, enabling designers to create products that better align with users’ expectations and thus enhance user satisfaction. Kansei engineering typically involves matching adjectives (emotional words) with visual imagery and using inferential calculations to identify the most suitable design solutions [22]. This process includes four steps: selection of visual imagery and adjectives, semantic space expansion, properties space expansion, and relationship modeling [22]. Semantic space extension refers to the rational screening, categorizing, and evaluating perceptual adjectives [23]. Semantic space expansion can usually be performed through factor analysis, cluster analysis, principal component analysis, or other methods [24]. Properties space expansion refers to the systematic definition and description of specific properties of visual images. The purpose of this process is to create a detailed properties space that enables features of visual images to be associated with semantic space. Relational model construction refers to establishing a mathematical or statistical model to describe and quantify the relationship between the attributes of the visual image (properties space) and the user’s emotional response (semantic space). Kansei engineering has been widely used in product development [24,25,26,27], user interface [28,29,30], and service design [31,32,33,34] and has achieved remarkable results. In the field of in-vehicle HMI, the dashboard’s visual imagery is the driver’s most frequently interacted element. For this reason, conducting a Kansei engineering study on the visual imagery of in-vehicle AR dashboards is necessary.

2.2. Multi-Criteria Decision Making (MCDM)

MCDM is a structured framework used to analyze decision problems with multiple complex objectives [35]. The core of MCDM lies in systematically analyzing and simplifying complex decision problems into manageable criteria and deriving the optimal decision through the trade-off of these criteria [36]. Handling uncertainty and subjectivity are common issues in the decision-making process, and MCDM methods provide decision makers with effective and robust decision support [37]. In practice, MCDM encompasses various methods, some of which include the analytic hierarchy process (AHP), grey relational analysis (GRA), the technique for order preference by similarity to the ideal solution (TOPSIS), and fuzzy TOPSIS [37]. These methods each offer unique advantages in addressing different types of decision problems and are suitable for various decision-making scenarios. The AHP, known for its structured approach and interpretability, is widely applied to various decision-making problems and remains one of the most commonly used MCDM methods to date [38]. Additionally, the AHP can be combined with triangular fuzzy numbers from fuzzy theory to make the decision-making process more realistic. However, when using the AHP or fuzzy AHP alone, the decision-maker judgment holds a dominant position within the hierarchy, which can lead to personal biases influencing the results [39]. GRA, on the other hand, is a flexible and adaptable MCDM method, but it is recommended to use weighted GRA, as it offers greater reliability and estimation accuracy compared to unweighted GRA [40]. TOPSIS is an MCDM method with relatively low computational complexity, making it suitable for handling large-scale decision problems. However, TOPSIS has certain limitations, such as the potential for rank reversal [41], and its use of Euclidean distance does not account for correlation, which may affect the results due to overlapping information [42]. Fuzzy TOPSIS, similar to the fuzzy AHP, incorporates triangular or trapezoidal fuzzy numbers to enhance decision-making accuracy.

In the field of Kansei engineering, evaluating visual imagery is a typical MCDM problem. For example, Jia and Tung [24] combined Kansei engineering with fuzzy theory to assess the visual imagery of wrist-worn wearable devices. Additionally, Lin and Zhai [26] applied TOPSIS within Kansei engineering to evaluate the visual imagery of automotive central touchscreens. In the realm of HMI, Li and Chen [28] conducted similar decision evaluations for visual imagery of waiting indicators. Wang and Yang [43] employed the GRA method to extract Kansei words related to wickerwork lamp products and conducted a study on Miryoku engineering. However, research combining MCDM with Kansei engineering in the automotive HMI domain remains relatively scarce. Although some scholars have used the AHP and GRA methods to evaluate the usability of automotive AR head-up displays (HUDs), they have not incorporated Kansei engineering methods [44]. In other research fields, fuzzy theory and TOPSIS have achieved certain successes in the study of visual imagery [24,26,28]. However, these methods are also limited by their respective theoretical foundations, and thus, they require a more comprehensive perspective. For instance, combining multiple MCDM methods can effectively address the limitations of individual methods, further enhancing the accuracy and reliability of decision analysis. Overall, MCDM methods have become important tools in modern decision analysis due to their scientific and systematic nature, with broad application prospects in complex systems or multi-criteria decision-making problems.

2.3. Fuzzy Analytic Hierarchy Process and Grey Relational Analysis (FAHP-GRA)

GRA is a significant method within MCDM. Its fundamental principle involves calculating grey relational degrees among variables to ascertain the degree of influence each factor has on the target variable, facilitating subsequent ranking and selection processes [45]. Compared to other decision-making methods, GRA exhibits clear advantages in handling uncertainty and fuzziness in MCDM processes [40]. In this study, the visual imagery investigation of the AR dashboard is evaluated based on the driver’s perception, and human subjective judgment is subjective and fuzzy, so it is particularly suitable for using the GRA analysis method. In practical applications, GRA has been widely employed in decision-making problems across various fields, such as engineering, management, and design. Its outstanding flexibility and effectiveness have been well demonstrated [46,47,48].

In addition, when conducting MCDM, we need to consider the weight value of each factor to achieve a more accurate and reliable assessment [49]. Specifically, when using GRA for decision making, weighted GRA is the optimal choice [40]. The fuzzy analytic hierarchy process (FAHP) [50] is a weight calculation method that combines fuzzy theory [51] and the analytic hierarchy process [52]. Because of the characteristics of human thinking and cognitive patterns in the actual decision-making process, the quantitative numerical approach may not accurately reflect the cognitive preferences of the decision maker [53]. If cognitive preferences are expressed through fuzzy semantic variables, they can provide a more flexible way of judgment [54]. Therefore, combining the FAHP and GRA methods can solve the standardized weighting problem inherent in the GRA model and promote the accuracy and science of MCDM assessment [55]. Additionally, the characteristics and advantages of these methods in MCDM have already been discussed earlier, and the FAHP-GRA method will be applied in the Kansei engineering process to construct and analyze relational models, providing decision support and design guidance for the visual imagery of in-vehicle AR dashboards.

2.4. Research Objectives

Interface evaluation is a typical MCDM problem. Additionally, the influence of HMI on drivers’ subjective preferences is complex and ambiguous. Therefore, this study employs a variety of rigorous analytical methods to conduct a comprehensive assessment of AR dashboard information design types. These methods include Kansei engineering, factor analysis, fuzzy theory, AHP, GRA, and factorial experiments. This study utilizes these objective research methods to review user perceptions and preferences regarding existing AR dashboard design types and conducts a design of experiments study on the main color, visual effects, and dial styling of AR dashboards. The objectives of this study are as follows:

• To establish evaluation dimensions and indicator weights for the visual imagery of AR dashboards.
• To rank the optimal design solutions for AR dashboards based on the visual imagery evaluation dimensions.
• To investigate the effects of three independent variables—main color, visual effects, and dial styling—on drivers’ preferences.
• To discuss the cross-validation results between drivers’ subjective evaluations and their visual imagery assessments of AR dashboards.

3. Methodological Procedures

The evaluation process for the AR dashboard HMI in this study is illustrated in Figure 1. Next, we will provide a detailed description of the three stages of Kansei engineering.

3.1. Phase 1: Selection and Expansion of Visual Imagery

The researchers collected 335 samples of dashboard interface design pictures and invited 12 in-vehicle HMI design experts to systematically analyze and discuss these picture samples. Among them were three user interface design experts, three user experience experts, three human factors researchers, and three product managers. Based on the discussion results of the in-vehicle HMI design experts, the researchers carried out the factorial experiment planning and HMI design for the AR dashboard. After systematic analysis, the main color, visual effect, and dial styling of the AR dashboard will be used as the independent variables in the factorial experiment. Previous studies have also indicated that the color and shape of a product are major factors influencing users’ emotional responses [56,57]. The visual effects and dial styling in this study are key aspects of the AR dashboard’s shape, making them highly relevant for studying the visual imagery of automotive dashboards. The main color and visual effect are a within-subjects factor, while the styling of the dial is a between-subjects factor; the main color is divided into three levels: blue (H:200, S:100, B:100), green (H:120, S:100, B:100), and yellow (H:60, S:100, B:100); the visual effect is divided into three levels: plane, vertical 3D, and horizontal 3D; and the styling of the dial is divided into two levels: round and unconventionally shaped. In a usability study of speedometers, Francois and Crave [58] noted that combination dials outperformed both analog and digital dials in the tasks of reading information and detecting dynamic speed changes. A combination dial is a design that uses both numeric and indicator elements to convey speed information. Therefore, we redesigned the AR dashboard interface of the SEV based on the speedometer design guidelines proposed by Francois and Crave [58]. The SEV-AR dashboard interface information in this experiment mainly consists of the speedometer and the power-to-weight ratio (PWR) dial. In designing the AR dashboard interface, we adhered to Nielsen’s principles of consistency, aesthetics, and minimalism [59]. The specific design proposal is shown in Figure 2. Based on the different levels of the three independent variables, we developed 18 AR dashboard interface design proposals. For example, the first design in the first row (Proposal 1) of Figure 2 features a blue planar and round dial.

The design proposals were formatted in a 12.3-inch (292.528mm × 109.698mm) format, and these designs were displayed on the liquid-crystal display. The evaluation task was carried out in a laboratory environment. We strictly followed the driver sight distance criterion proposed by Dreyfuss and Associates [60], where participants were asked to sit at a distance of 550 mm from the sight distance of the dashboard screen. While observing the design proposals, participants were able to swipe left and right to view each design proposal while completing the scale questionnaire. The scoring was on a 7-point Likert scale (1 for very low, 4 for average, 7 for very high). Figure 3 illustrates the process by which participants switched between different design schemes during the experiment.

3.2. Phase 2: Selection and Expansion of Adjectives

Visual imagery adjectives can effectively respond to the user’s mental feelings [28]. For instance, shapes and colors within visual imagery can have different impacts on users’ psychological responses [56]. One of the key tasks for designers is to evoke specific emotional responses from users by manipulating visual imagery elements such as shapes and colors [57]. Therefore, controlling the visual imagery of AR dashboards is a critical method for designers to convey information to drivers and elicit emotional responses. At this stage, we first collected many adjectives related to the visual imagery of the dashboard from automotive portals, design resource websites, and the relevant research literature. For example, we extracted adjectives from user reviews of dashboard HMIs on automotive portals and design resource websites. Subsequently, after expert focus group discussions, adjectives unsuitable for describing the in-vehicle HMI were eliminated, leaving 130 adjectives for subsequent experiments. In the following study phase, we invited 12 designers and researchers related to the vehicle HMI to participate in the experiment. The participants were asked to select 40 to 50 adjectives from the aforementioned 130 that best describe the AR dashboard interface. Finally, the researchers selected the 40 most recognized adjectives based on the frequency of votes cast by the participants for the study of the factor analysis scale.

Factor analysis is a statistical method that uses a system of indicators to analyze or measure the extent to which multiple factors have an impression on an objective phenomenon [61]. Factor analysis is the most used analysis method in Kansei engineering, which can extract key perceptual factors from a large number of sensibility words, and these factors can be used to guide the subsequent design. Recently, many scholars have demonstrated that factor analysis is a scientific and reliable method for studying visual imagery [24,28,62]. Therefore, we used factor analysis in this phase to extract imagery adjectives for the AR dashboard interface. Specifically, participants were invited to experience the AR dashboard design sample from Phase 1 and then asked to evaluate 40 imagery adjectives using a 5-point Likert scale (1 for very inappropriate, 3 for average, 5 for very appropriate). The collected data will be factor analyzed to extract the adjectives that match the AR dashboard interface.

3.3. Phase 3a: Relationship Modeling—Fuzzy Analytic Hierarchy Process (FAHP) to Determine Visual Imagery Evaluation Dimension Weights

In this stage, FAHP weights are calculated for the adjectives (assessment dimensions) derived from the factor analysis, and the specific calculation steps are as follows.

Step 1: Perform a pairwise comparison of the assessment dimensions.

This study will invite user interface designers, user experience designers, and product managers to form an evaluation team to compare the importance of visual image dimensions in pairs. The measurement scale uses a semantic scale with a 1–9 level pairwise comparison [52], which is then converted into a triangular fuzzy number [50,63], as shown in Table 1 and Figure 4.

Table 1. Triangular fuzzy conversion scale.

Linguistic Scale	AHP Scale	Triangular Fuzzy Number Scale
		Left Endpoint	Middle Value	Right Endpoint
Equal importance	1	1	1	3
Slight importance	3	1	3	5
Important	5	3	5	7
Strong importance	7	5	7	9
Extreme importance	9	7	9	9
Slight unimportance	1/3	1/5	1/3	1
Unimportant	1/5	1/7	1/5	1/3
Strong unimportance	1/7	1/9	1/7	1/5
Extreme unimportance	1/9	1/9	1/9	1/7

Table 1. Triangular fuzzy conversion scale.

Linguistic Scale	AHP Scale	Triangular Fuzzy Number Scale
		Left Endpoint	Middle Value	Right Endpoint
Equal importance	1	1	1	3
Slight importance	3	1	3	5
Important	5	3	5	7
Strong importance	7	5	7	9
Extreme importance	9	7	9	9
Slight unimportance	1/3	1/5	1/3	1
Unimportant	1/5	1/7	1/5	1/3
Strong unimportance	1/7	1/9	1/7	1/5
Extreme unimportance	1/9	1/9	1/9	1/7

Figure 4. Linguistic variables describing weights of the FAHP.

Figure 4. Linguistic variables describing weights of the FAHP.

Step 2: Create a pairwise comparison matrix.

The values of the pairwise comparison results of n image adjective dimensions are placed in the upper triangular part of the pairwise comparison matrix A, and the lower triangular part is the reciprocal of the relative position, that is, a_ji = 1/a_ij. Matrix A can be expressed as follows:

$$\mathrm{A}=\left[\begin{array}{ccc}\begin{array}{cc}1& a_{12}\end{array}& \cdots & a_{1n}\\ \begin{array}{cc}{\displaystyle \frac{1}{a_{12}}}& 1\end{array}& \cdots & a_{2n}\\ \begin{array}{cc}\begin{array}{c}⋮\\ {\displaystyle \frac{1}{a_{1n}}}\end{array}& \begin{array}{c}⋮\\ {\displaystyle \frac{1}{a_{2n}}}\end{array}\end{array}& \begin{array}{c}\ddots \\ \cdots \end{array}& \begin{array}{c}⋮\\ 1\end{array}\end{array}\right]$$

(1)

Step 3: Calculate the maximum eigenvalue and conduct consistency identification.

$$\overline{W_i}=\sqrt[n]{\prod _{i=1}^n{A}_{ij}}$$

(2)

$$W_i=\overline{W_i}/\sum _{i=1}^n{W}_i$$

(3)

Next, the maximum eigenvalue λ_max is found based on W_i and the comparison matrix A, as shown in Formula (4). Finally, find the random consistency ratio CR required in this step. When the CR value is not greater than 0.1, the importance matrix has satisfactory consistency. CI is the consistency index, and RI is the average random consistency index. The value range of the RI is visible in

Table 2. Value of random indexes (RI).

Matrix Size (n)	3	4	5	6	7
RI value	0.58	0.90	1.12	1.24	1.32

.

$${\lambda }_{max}={\displaystyle \frac{1}{n}}/\sum _{i=1}^n{\displaystyle \frac{{\left(AW\right)}_i}{W_i}}$$

(4)

$$CI={\displaystyle \frac{{\lambda }_{max}-n}{n-1}}$$

(5)

$$CR=CI/RI$$

(6)

Step 4: Convert the original scores into triangular fuzzy numbers and establish a fuzzy pairwise comparison matrix.

After passing the consistency test, each internal value of the pairwise comparison matrix A is converted into a triangular fuzzy number ${\tilde{M}}_{ij}$, and a fuzzy pairwise comparison matrix M is established. That is, ${\tilde{M}}_{ij}=(L_{ij},M_{ij},R_{ij})$, the fuzzy number of the evaluation dimension i is relative to the evaluation dimension j, and the lower triangular part of the matrix M is ${\tilde{M}}_{ji}=1/{\tilde{M}}_{ij}$. The matrix M can be expressed as follows:

$$\mathrm{M}=\left[{\tilde{M}}_{ji}\right]=\left[\begin{array}{ccc}\begin{array}{cc}1,1,1& {\tilde{M}}_{12}=(L_{12},M_{12},R_{12})\end{array}& \cdots & {\tilde{M}}_{1j}=(L_{1j},M_{1j},R_{1j})\\ \begin{array}{cc}{\tilde{M}}_{21}=1/{\tilde{M}}_{12}& \mathrm{1,1},1\end{array}& \cdots & {\tilde{M}}_{2j}=(L_{2j},M_{2j},R_{2j})\\ \begin{array}{cc}\begin{array}{c}⋮\\ {\tilde{M}}_{j1}=1/{\tilde{M}}_{1j}\end{array}& \begin{array}{c}⋮\\ {\tilde{M}}_{j2}=1/{\tilde{M}}_{2j}\end{array}\end{array}& \begin{array}{c}\ddots \\ \cdots \end{array}& \begin{array}{c}⋮\\ 1,1,1\end{array}\end{array}\right]$$

(7)

Step 5: Calculate triangular fuzzy numbers and fuzzy weights.

Perform a geometric mean operation on the values in the fuzzy pairwise comparison matrix M to obtain the geometric mean triangular fuzzy number $\stackrel{`}{M}=(\stackrel{`}{L_i},\stackrel{`}{M_i},\stackrel{`}{R_i})$ in each column of each evaluation dimension, and then add up the geometric mean triangular fuzzy numbers in each column. At the same time, to ensure that the left boundary value of the triangular fuzzyweight issmaller than the right boundary value, the summed triangular fuzzy number needs to be reciprocally converted. That is, $\stackrel{`}{\stackrel{`}{M}}=(\stackrel{`}{{\stackrel{`}{L}}_i},\stackrel{`}{{\stackrel{`}{M}}_i},\stackrel{`}{{\stackrel{`}{R}}_i})=(1/\stackrel{`}{{\stackrel{`}{R}}_i},1/\stackrel{`}{{\stackrel{`}{M}}_i},1/\stackrel{`}{{\stackrel{`}{L}}_i},)$. Finally, the triangular blur weight $\widetilde{W_i}=(1/\stackrel{`}{{\stackrel{`}{R}}_i}{L}_i,1/\stackrel{`}{{\stackrel{`}{M}}_i}{M}_i,1/\stackrel{`}{{\stackrel{`}{L}}_i}{R}_i)$ is calculated. The weight calculation methods of the FAHP and AHP are similar and can be deduced concerning Formulas (2) and (3), which will not be described again.

Step 6: Defuzzification and normalization.

Defuzzification is performed on the obtained triangular fuzzy weight $\widetilde{W_i}$, which is, converted into a real value $D{W}_i$. In addition, normalization needs to be performed again to make the sum of the importance of each evaluation dimension equal to 1. Finally, the fuzzy weight value $D\stackrel{`}{W_i}$ of each element is obtained. Assuming the triangular fuzzy weight $\widetilde{W_i}=(W{L}_i,W{M}_i,W{R}_i)$, the defuzzification and normalization formulas are as follows:

$${DW}_i={\displaystyle \frac{(W{R}_i-W{L}_i)+(W{M}_i-W{L}_i)}{3}}+W{L}_i$$

(8)

$$D\stackrel{`}{W_i}={\displaystyle \frac{{DW}_i}{\sum _{i=1}^n{DW}_i}}$$

(9)

3.4. Phase 3b: Relationship Modeling—Evaluating Visual Imagery Using FAHP-GRA

In this stage, a second user questionnaire was constructed based on the design proposals and adjectives derived from Phases 1 and 2, thereby measuring the driver’s visual image evaluation of the AR instrument panel design proposal. The questionnaire adopts a 7-point Likert scale (1 for very low, 4 for medium, and 7 for very high). Subsequently, the questionnaire data were calculated by FAHP-GRA to obtain the scores of each AR dashboard design proposal. The calculation steps of FAHP-GRA are as follows.

Step 1: Construct reference and comparison sequences.

Based on the final scores of the 18 AR dashboards, the optimal value in each evaluation dimension is selected as the reference sequence C₀. At the same time, the scores of the 18 AR dashboards will be used as the comparison sequences C₁, C₂, C₃, …, C₁₈.

Step 2: Perform non-dimensionalization.

Although all assessment dimensions use a 7-point Likert scale, differences in the resulting data range may lead to numerical instability or calculation accuracy issues. Therefore, the non-dimensionalization of reference and comparison sequences is required to improve the stability of data calculations.

$$X_i(k)={\displaystyle \frac{C_i\left(k\right)}{C_k}}$$

(10)

Among them, $C_k=1/(\mathrm{n}+1){\sum }_{i=0}^n{C}_i\left(k\right),k=1,2,\dots ,m.$

Step 3: Determine the optimal value range of the distinguishing coefficient ρ.

Before performing the gray correlation calculation, the distinguishing coefficient ρ value must be determined. The value range of ρ is (0, 1). Usually, ρ = 0.5 is taken. However, since the distinguishing coefficient will affect the arrangement of related sequences, we should not simply apply ρ = 0.5 or other values. Necessary calculations need to be performed to determine the ρ value [64]. Therefore, this step adopts the calculation formula for the value range of the distinguishing coefficient ρ proposed by Guo and Guo [65] as follows:

$${\mathsf{\rho }}_1={\displaystyle \frac{{\mathrm{⊿}}_{0i}\left(k\right)}{{⊿}_{max}}}·{\displaystyle \frac{1}{e-1}}$$

(11)

$${\mathsf{\rho }}_2=(e-1)·{\mathsf{\rho }}_1$$

(12)

For this gray relational system, the distinguishing coefficient value is optimal between [ρ1, ρ2].

Step 4: Calculate the grey relational degree of each sequence.

After determining the ρ value, the grey relational degree can be calculated [66]. The formula is as follows:

$${\xi }^{\left(i\right)}(\mathrm{k})={\displaystyle \frac{\underset{i}{\min }\underset{k}{\min }\left|X^{\left(0\right)}\left(k\right)-\left.X^{\left(i\right)}\left(k\right)\right|+\mathsf{\rho } \underset{i}{\max }\underset{k}{\max }\left|X^{\left(0\right)}\left(k\right)-\right.\left.X^{\left(i\right)}\left(k\right)\right|\right.}{\left|X^{\left(0\right)}\left(k\right)-\right.\left.X^{\left(i\right)}\left(k\right)\right|+\mathsf{\rho } \underset{i}{\max }\underset{k}{\max }\left|X^{\left(0\right)}\left(k\right)-\right.\left.X^{\left(i\right)}\left(k\right)\right|}}$$

(13)

Step 5: Calculate the weighting grey relational degree.

To compare different AR dashboard design proposals more scientifically and comprehensively, it is necessary to integrate the weight values and grey relational degree of each evaluation dimension [67]. Let the grey relational degree of each design proposal be γⁱ, and its formula is as follows:

$${\mathsf{\gamma }}^{\left(i\right)}=\sum _{k=1}^n{W}_i{·\mathsf{\gamma }}^{\left(i\right)}(\mathrm{k})$$

(14)

When the γⁱ value is larger, it indicates that the visual image of the AR dashboard design proposal is better.

4. Analysis and Results

4.1. Visual Imagery Adjective Extraction Results

After referring to the HMI design of AR dashboards, the expert team sifted and sorted out the 40 most common adjectives for evaluating AR dashboard interfaces from 130 imagery adjectives, which are all positive (see

Table 3. Forty most common adjectives for visual imagery.

Visual Imagery Adjectives
Detailed	Dynamic	Rich	Concise	Vivid
Responsive	Cool	Intelligent	Technological	Immersive
Innovative	Secure	Efficient	Futuristic	Aesthetic
Unique	Stunning	Abstract	Visual	Premium
Reliable	Smooth	Gorgeous	Personalized	Attractive
Harmonious	Natural	Interesting	Diverse	Practical
Orderly	Dreamlike	Clear	Sleek	Intuitive
Trustworthy	Grand	Rhythmic	Fresh	Agile

).

Next, we combined 40 adjectives with 18 AR dashboard HMI design proposals. Using purposive sampling, a total of 140 questionnaires were collected. All participants were required to have a driver’s license and possess certain driving information recognition capabilities. In the end, 123 valid questionnaires were obtained, 60 from males and 63 from females, with and an average age of 28.61 (SD = 5.46). Subsequently, we conducted the first factor analysis on the questionnaire data and a second factor analysis on the 23 adjectives with factor loadings greater than 0.6 (see

Table 4. The 23 adjectives with factor loadings higher than 0.6.

Adjectives	Initial	Extraction	Adjectives	Initial	Extraction
Premium	1.000	0.803	Efficient	1.000	0.722
Natural	1.000	0.762	Reliable	1.000	0.715
Unique	1.000	0.758	Intelligent	1.000	0.706
Aesthetic	1.000	0.758	Trustworthy	1.000	0.704
Attractive	1.000	0.751	Concise	1.000	0.701
Personalized	1.000	0.741	Diverse	1.000	0.699
Dreamlike	1.000	0.738	Stunning	1.000	0.679
Visual	1.000	0.736	Interesting	1.000	0.677
Smooth	1.000	0.735	Responsive	1.000	0.661
Gorgeous	1.000	0.734	Innovative	1.000	0.646
Technological	1.000	0.728	Rich	1.000	0.605
Clear	1.000	0.726

Extraction method: principal component analysis.

).

As shown in

Table 5. The KMO and Bartlett test results.

Visual Imagery Adjectives
Kaiser Meyer Olkin		0.896
Bartlett’s Sphericity Test	Approximate Chi-Squared	1949.631
	df	253
	p	<0.001 ***

* Significantly different at α = 0.05 level (p < 0.05). ** Significantly different at α = 0.01 level (p < 0.01). *** Significantly different at α = 0.001 level (p < 0.001).

, after the second factor analysis, KMO = 0.896, Bartlett = 1949.631, and p < 0.001 (df = 253), indicating a statistically significant difference. This result suggests that the correlation matrices in the original group have common factors and are suitable for factor analysis.

In addition, according to the principal component method and the eigenvalue principle, there are five factors with an eigenvalue greater than 1, and their total explained variance is 71.675% (see

Table 6. Total variance explained.

Component	Initial Eigenvalues			Squares Loading Extraction			Transformed Squares Loading
	Total	Variance (%)	Accumulative (%)	Total	Variance (%)	Accumulative (%)	Total	Variance (%)	Accumulative (%)
1	10.313	44.837	44.837	10.313	44.837	44.837	6.103	26.533	26.533
2	2.573	11.189	56.026	2.573	11.189	56.026	3.353	14.579	41.112
3	1.301	5.657	61.683	1.301	5.657	61.683	2.801	12.180	53.292
4	1.186	5.158	66.840	1.186	5.158	66.840	2.500	10.868	64.160
5	1.112	4.834	71.675	1.112	4.834	71.675	1.728	7.515	71.675

). Generally speaking, a value higher than 70% indicates a good level of explanation. Therefore, five groups of similar factors were extracted at this stage, as shown in Figure 5.

In

Table 7. The transformed component matrices.

Adjectives	Component
	Factor 1	Factor 2	Factor 3	Factor 4	Factor 5
Unique	0.841	0.157	0.132	0.022	0.093
Dreamlike	0.829	0.097	0.070	0.139	0.129
Gorgeous	0.781	0.198	−0.060	0.279	0.058
Innovative	0.767	0.161	0.136	0.032	0.110
Diverse	0.755	0.228	0.184	0.200	0.054
Stunning	0.751	0.200	0.028	0.195	0.190
Personalized	0.738	0.376	0.189	0.126	0.056
Interesting	0.704	0.356	0.046	0.057	0.222
Rich	0.612	0.309	0.195	0.312	0.012
Premium	0.260	0.755	0.054	0.320	0.248
Aesthetic	0.269	0.701	0.157	0.330	0.246
Attractive	0.312	0.698	0.236	0.322	0.084
Technological	0.403	0.692	0.262	−0.133	−0.018
Intelligent	0.428	0.673	0.225	0.117	0.069
Visual	0.118	0.334	0.774	0.056	−0.090
Efficient	0.212	0.078	0.740	0.315	0.157
Clear	0.201	0.098	0.730	0.370	0.084
Concise	−0.053	0.155	0.716	0.029	0.400
Trustworthy	0.047	0.227	0.285	0.743	0.132
Responsive	0.329	0.205	0.078	0.709	−0.049
Reliable	0.254	0.085	0.310	0.672	0.310
Natural	0.264	0.101	0.103	0.021	0.819
Smooth	0.175	0.238	0.242	0.339	0.689

Extraction method: principal component analysis. Transformed method: Kaiser normalized maximum variance method.

, the difference between each adjective in the five components is obvious, and it does not overwhelm multiple components. Moreover, the loadings of the factors are all higher than the excellent standard of 0.6, indicating that these adjectives have very high structural validity for evaluating the design of the AR dashboard HMI.

After factor analysis, twenty-three adjectives in five groups were obtained in this stage. Since the adjectives in each group are related, we invited language and literature experts to rename each group. The results are shown in

Table 8. Renamed adjectives and their codes.

Factor	Adjective Groups	Factor Renaming	Code
1	Unique, Dreamlike, Gorgeous, Innovative, Diverse, Stunning, Personalized, Interesting, Rich	Novel and Splendid	N&S
2	Premium, Aesthetic, Attractive, Technological, Intelligent	Technological and Aesthetic	T&A
3	Visual, Efficient, Clear, Concise	Visible and Concise	V&C
4	Trustworthy, Responsive, Reliable	Agile and Reliable	A&R
5	Natural, Smooth	Smooth and Natural	S&N

. These results created five basic visual image evaluation dimensions for the next stage of research.

4.2. Visual Imagery Weighting Results

At this stage, 12 experts were invited to rate important pairs of evaluation dimensions. Experts include in-vehicle HMI user interface (UI) designers, user experience (UE) designers, ergonomics researchers (ERs), and product managers (PMs). Next, we performed an AHP weight calculation (referring to Formulas (2) and (3)) and a consistency test (referring to Formulas (4), (5), and (6)) with the ratings of the 12 experts.

In

Table 9. AHP weights and consistency test results of experts.

Expert Code	N&S	T&A	V&C	A&R	S&N	CI	CR
UI 1	0.049	0.100	0.443	0.193	0.214	0.050	0.045
UI 2	0.050	0.050	0.367	0.379	0.155	0.070	0.063
UI 3	0.056	0.485	0.255	0.107	0.097	0.086	0.076
UE 1	0.043	0.071	0.057	0.335	0.494	0.076	0.068
UE 2	0.073	0.061	0.172	0.394	0.300	0.062	0.055
UE 3	0.055	0.173	0.239	0.318	0.216	0.039	0.035
ER 1	0.039	0.094	0.567	0.216	0.084	0.071	0.064
ER 2	0.046	0.078	0.228	0.206	0.441	0.072	0.064
ER 3	0.129	0.245	0.499	0.083	0.045	0.069	0.062
PM 1	0.037	0.171	0.439	0.182	0.171	0.026	0.024
PM 2	0.112	0.192	0.192	0.239	0.265	0.081	0.072
PM 3	0.046	0.079	0.166	0.355	0.355	0.060	0.053

, the CR values of all expert ratings are less than 0.1; that is, the weight matrix of each expert has satisfactory consistency. Therefore, the importance of weight calculation, defuzzification (refer to Formula (8)), and regularization (refer to Formula (9)) of the FAHP are performed again, and the results are shown in

Table 10. FAHP weights of experts.

Expert Code	N&S	T&A	V&C	A&R	S&N
UI 1	0.055	0.123	0.403	0.218	0.201
UI 2	0.056	0.048	0.389	0.341	0.166
UI 3	0.071	0.449	0.264	0.126	0.090
UX 1	0.052	0.078	0.050	0.343	0.477
UX 2	0.092	0.059	0.191	0.395	0.264
UX 3	0.062	0.217	0.252	0.302	0.168
ER 1	0.046	0.116	0.536	0.217	0.085
ER 2	0.050	0.084	0.257	0.198	0.411
ER 3	0.137	0.254	0.460	0.097	0.052
PM 1	0.036	0.213	0.408	0.191	0.153
PM 2	0.145	0.231	0.195	0.221	0.208
PM 3	0.050	0.085	0.187	0.366	0.312
Geometric mean	0.065	0.132	0.260	0.232	0.180
Normalized weight	0.074	0.152	0.299	0.268	0.207

.

The results show that the weight of novel and splendid (N&S) is 0.074, technological and aesthetic (T&A) is 0.152, visible and concise (V&C) is 0.299, agile and reliable (A&R) is 0.268, and smooth and natural (S&N) is 0.207.

4.3. FAHP-GRA Calculation Results

At this stage, 100 drivers (60 males and 40 females) were invited to evaluate the five dimensions of N&S, T&A, V&C, A&R, and S&N of 18 AR instrument panels. Their average age was 29.76 years (SD = 5.03). The evaluation score results are processed according to the average, and the AR dashboard design solution’s performance value and comparison sequence are obtained (see

Table 11. Performance scores of 18 proposals.

Proposal Codes	N&S	T&A	V&C	A&R	S&N
Proposal 1	4.380	4.500	5.380	4.800	4.980
Proposal 2	5.220	5.260	4.840	4.600	4.740
Proposal 3	4.740	4.580	4.700	4.580	4.480
Proposal 4	4.380	4.300	4.600	4.420	4.460
Proposal 5	4.460	4.420	4.340	4.280	4.420
Proposal 6	4.260	4.220	4.160	4.060	4.060
Proposal 7	3.900	3.880	4.040	3.840	3.840
Proposal 8	4.080	3.920	3.860	3.880	3.900
Proposal 9	3.860	3.840	3.960	3.940	3.920
Proposal 10	4.400	4.680	5.340	4.820	4.980
Proposal 11	4.820	4.820	4.440	4.220	4.260
Proposal 12	4.900	4.780	4.800	4.440	4.460
Proposal 13	4.300	4.280	4.760	4.480	4.560
Proposal 14	4.440	4.280	4.340	4.180	4.100
Proposal 15	4.840	4.640	4.620	4.400	4.280
Proposal 16	4.160	4.000	4.520	4.120	4.140
Proposal 17	4.220	4.060	3.920	3.740	3.580
Proposal 18	4.360	4.120	4.360	3.860	4.040

).

Table 11 shows the reference sequence $C_0=\left[\begin{array}{ccc}5.220& 5.260& \begin{array}{ccc}5.380& 4.820& 4.980\end{array}\end{array}\right]$ of the AR dashboard and the comparison sequence from Proposals 1 to 18. First, we performed non-dimensionalization processing on the reference and comparison sequences according to Formula (10). Next, we calculated the values of the distinguishing coefficients ρ1 and ρ2 according to Formulas (11) and (12). The results show that the distinguishing coefficient is optimal between 0.300 and 0.516. In this study, the ρ value takes the intermediate value of 0.4 for subsequent calculations. Finally, we calculated the gray relational degree of each proposal according to Formula (13). We substituted its gray relational degree and the FAHP results into Formula (14) to obtain the overall relational degree γⁱ (see

Table 12. Weighted relational degree of each proposal.

Proposal Codes	N&S	T&A	V&C	A&R	S&N	γi	Sequence
Proposal 1	0.031	0.066	0.299	0.259	0.207	0.862	2
Proposal 2	0.074	0.152	0.158	0.194	0.146	0.725	3
Proposal 3	0.041	0.071	0.141	0.189	0.111	0.553	5
Proposal 4	0.031	0.058	0.131	0.158	0.109	0.487	8
Proposal 5	0.033	0.063	0.110	0.138	0.105	0.449	10
Proposal 6	0.028	0.055	0.099	0.115	0.080	0.378	13
Proposal 7	0.023	0.046	0.093	0.099	0.070	0.331	17
Proposal 8	0.025	0.046	0.085	0.102	0.072	0.331	16
Proposal 9	0.023	0.045	0.090	0.106	0.073	0.336	15
Proposal 10	0.031	0.077	0.281	0.268	0.207	0.863	1
Proposal 11	0.044	0.087	0.117	0.131	0.092	0.472	9
Proposal 12	0.048	0.084	0.153	0.161	0.109	0.556	4
Proposal 13	0.029	0.057	0.148	0.168	0.120	0.523	6
Proposal 14	0.032	0.057	0.110	0.127	0.082	0.409	11
Proposal 15	0.045	0.074	0.133	0.155	0.094	0.501	7
Proposal 16	0.027	0.048	0.124	0.121	0.085	0.404	12
Proposal 17	0.028	0.050	0.088	0.093	0.061	0.319	18
Proposal 18	0.030	0.052	0.112	0.100	0.079	0.373	14

).

The closer the value of γⁱ is to 1, the closer the proposal is to the ideal proposal (reference sequence). In Table 12, the blue planar and unconventional-shaped dial design (Proposal 10) has the closest value of γⁱ to 1 (γ¹⁰ = 0.863). Therefore, this proposal is the best AR dashboard design.

4.4. Factorial Experiment Results

This experiment utilizes a 3 (main color) × 3 (visual effect) × 2 (dial styling) mixed factorial design. During the experiment, drivers were asked to rate both the visual imagery of the AR dashboard proposal and their subjective preferences. We conducted a three-way ANOVA on the subjective preference outcome data. Additionally, LSD post hoc test analysis was performed on statistically significant variables. It is worth noting that during the ANOVA analysis of repeated measures data, it is necessary to first examine the sphericity of the data [68]. In this study, Mauchly’s test of sphericity indicated a significant difference (p < 0.05), and we applied the Greenhouse Geisser correction to adjust the degrees of freedom [69,70]. The results are shown in

Table 13. The mixed factorial ANOVA results of subjective preference (after correction).

Source	SS	df	MS	F	p	η2	Post Hoc
Main color	100.287	1.704	58.847	30.710	<0.001 ***	0.239	Blue > Green > Yellow
Visual effect	0.507	1.819	0.279	0.154	0.838	0.002
Dial styling	4.551	1.000	4.551	0.549	0.461	0.006
Main color × visual effect	7.307	3.484	2.097	3.426	0.013 *	0.034
Main color × dial styling	5.016	1.704	2.943	1.536	0.220	0.015
Visual effect × dial styling	10.702	1.819	5.883	3.249	0.046 *	0.032
Main color × visual effect × dial styling	1.711	3.484	0.491	0.802	0.509	0.008

* Significantly different at α = 0.05 level (p < 0.05). ** Significantly different at α = 0.01 level (p < 0.01). *** Significantly different at α = 0.001 level (p < 0.001).

.

The main color has a significant main effect (F_1.704,98 = 30.710, p < 0.001; η² = 0.239). Further analysis using the LSD post hoc test showed significant differences in subjective preferences between blue (M = 4.587, SE = 0.096), green (M = 4.143, SE = 0.110), and yellow (M = 3.770, SE = 0.131). The subjective preference for blue is significantly higher than for green and yellow, while the subjective preference for green is also significantly higher than for yellow. However, the main effect of visual effects was not significant (F_1.819,98 = 0.154, p = 0.838 > 0.05; η² = 0.002). Similarly, dial styling has no significant main effect (F_1.000,98 = 0.549, p = 0.461 > 0.05; η² = 0.006).

The interaction between the main color and the visual effect was significant (F_3.484,98 = 3.426, p = 0.013 < 0.05; η² = 0.034). To test the differences between groups within a certain level of an independent variable [71], we conducted a simple effects analysis. A simple effects analysis on the main color reveals that only the visual effect of green is significantly different (see Figure 6). Specifically, when the main color is green, the visual effect of vertical 3D is significantly worse than that of flat and horizontal 3D.

Additionally, there is a significant interaction between visual effects and dial styling (F_1.819,98 = 3.249, p = 0.046 < 0.05; η² = 0.032). In the simple effects analysis of dial styling, significant differences were found in the visual effects of different dial styles. For round dials, the visual effect of vertical 3D is significantly better than horizontal 3D. For unconventional-shaped dials, the visual effect of horizontal 3D is significantly better than vertical 3D (see Figure 7).

Overall, the main color (η² = 0.239) has a greater impact on subjective preference than dial styling (η² = 0.006) and visual effects (η² = 0.002). In comparison, dial styling impacts subjective preference more than visual effects. According to the results of FAHP-GRA, the top five design solutions are all blue, with the highest-ranked design solution being a blue planar and unconventional-shaped dial design (Proposal 10).

5. Discussion

5.1. Discussion of the Results

This study discusses the effects of the main color, visual effect, and dial styling of AR dashboards on drivers’ visual imagery evaluation and subjective preference. Previous studies have shown that in in-vehicle AR user interfaces, blue, green, and yellow colors exhibit superior robustness and response efficiency compared to other colors [72]. This study further concluded through visual imagery experiments that among the main colors, drivers prefer blue the most, followed by green, with yellow being the least preferred. However, recent studies have shown that AR heads-up displays (HUDs) with green as the main color have the shortest response time [12]. In addition, there are studies showing that red, yellow, green, and orange perform better than other colors in terms of visual search performance and cognitive efficiency [14]. These studies have inconsistent evaluation results regarding dominant colors, possibly due to experimental factors such as display technology [73], ambient illumination [13], and driving scenes [14]. For example, the photophysical properties of blue luminescent materials are worse than those of other colors [74], particularly in terms of luminous efficiency, maximum brightness, and the working life of blue quantum dots [75]. In any case, previous research mostly focused on the objective effectiveness of in-vehicle AR display technology and did not study the impact of in-vehicle AR interface displays on the driver’s emotional experience. This research has identified a new direction for the automotive AR interface display field. From the dimensions of the driver’s emotional experience, blue is the best choice. Although blue light display technology is more challenging to develop than other colors of light, we recommend that automobile manufacturers and related technicians prioritize advancing blue light display technology to meet the emotional needs of most drivers.

In addition, the interaction between visual effects and dial styling was significantly different. This is similar to the results of Chen and Lu [76], which showed no significant difference between round and unconventional-shaped (hexagonal) designs in balanced aesthetics experiments, but there is a significant difference between vertical and horizontal ellipse images. Our results indicate that there is a significant difference between the vertical 3D design of a circle and the horizontal 3D design of a round. The round vertical 3D design was more popular among drivers, and the visual imagery of this proposal was rated higher. Our study further revealed an interaction between round and unconventional-shapes in terms of visual effect, where the vertical 3D design of unconventional-shapes was significantly less preferred than the horizontal 3D design of unconventional-shapes. Drivers preferred the unconventional-shaped horizontal 3D design more, and the visual imagery of that proposal was rated higher. This finding complements Chen and Lu [76]’s theory on the interaction between styling and visual effect in visual aesthetics. Research on the emotional aspects of products suggests that emotional experiences will help increase product utilization and influence future purchase choices [77]. Therefore, in developing in-vehicle AR dashboard HMIs, automobile manufacturers should pay attention to the effects of the main color, visual effect, and dial styling on the driver’s emotional experience.

5.2. Methodological Contributions

One of the important innovations of this study is the combination of various MCDM and Kansei engineering methods to evaluate the visual image of vehicle HMIs. The visual image evaluation of vehicle HMIs is influenced by factors such as participants’ personal preferences, culture, and knowledge level and exhibits typical gray system characteristics and ambiguity [78]. Therefore, the decision-making method using FAHP-GRA is particularly suitable. Recently, Yunuo and Xia [44] confirmed in a study on vehicle-mounted AR-HUD that AHP-GRA is more reliable than entropy weights TOPSIS in determining weights. Although the study by Yunuo and Xia [44] described the implementation of AHP-GRA in HMIs in great detail, it did not use the methods of Kansei engineering and factor analysis to construct evaluation indicators, which may affect the objectivity of the program. In previous research, we made a preliminary attempt to apply AHP-GRA in the usability evaluation of application programming interfaces and determined the feasibility of this method in the HMI field through a triangulation model [79]. This study further combines fuzzy system theory with the AHP-GRA method and introduces the research method of Kansei engineering to conduct visual image research on vehicle-mounted AR instrument panels. Additionally, we conducted interactive verification of the subjective preference results and visual image assessment results through factorial experiments to ensure the reliability of the MCDM results.

5.3. Limitations and Future Directions

Despite the rigorous examination conducted in this study, several limitations should be considered. First, this research only explored the visual imagery of three variables within AR dashboards. Future studies could expand to include more variables, such as the size, layout, and brightness of elements in AR dashboards. Second, the participants in this study were primarily drivers from the Chinese region, so caution should be exercised when generalizing the findings to consumers in other countries or regions. Future research could involve a comparative analysis of drivers from different countries or regions. Finally, the results of this study have not yet been tested in real-world settings. Future research may need to conduct usability tests and incorporate eye-tracking data and driving performance into the evaluation.

6. Conclusions

This study is highly innovative, both from a methodological perspective and within the context of in-vehicle HMI research. Methodologically, this study integrates Kansei engineering, fuzzy system theory, AHP, and GRA, proposing a subjective evaluation method and process for assessing the visual imagery of in-vehicle HMIs. This approach helps reduce the uncertainty in HMI design and effectively addresses the ambiguity inherent in human factors. The FAHP results reveal that the dimensions affecting the visual imagery evaluation of AR dashboards include novelty and splendor, technological and aesthetic aspects, visibility and conciseness, agility and reliability, and smoothness and naturalness. Among these dimensions, “visibility and conciseness” received the highest weight, while “novelty and splendor” received the lowest. Further GRA analysis indicated that the design featuring blue planar and unconventional-shaped dials (Proposal 10) was the optimal choice based on visual imagery criteria. Conversely, the design with yellow vertical 3D and unconventional-shaped dials (Proposal 17) was the least favored, a finding that was corroborated by factorial experiments.

The factorial experiment results demonstrated that the main color of the AR dashboard had the most significant impact on drivers’ subjective preferences. Blue was the most favored main color, followed by green, with yellow being the least favored. Avoiding vertical 3D visual effects in green AR dashboards is also recommended. The interaction between visual effects and dial styling in AR dashboards also requires special attention. In round dials, drivers preferred vertical 3D effects more than horizontal 3D effects. Conversely, horizontal 3D effects were more favored in unconventional-shaped dials than vertical 3D effects. These findings provide scientific and detailed guidance for future SEV-AR dashboard HMI designs, helping to enhance the in-vehicle user experience and, consequently, improving SEV vehicles’ market competitiveness.