Optimization of Underground Cavern Sign Group Layout Using Eye-Tracking Technology

Abstract

Efficient sign layouts play a crucial role in guiding driving in underground construction caverns and enhancing transportation safety. Previous studies have primarily focused on evaluating drivers’ gaze behavior in tunnels to optimize individual traffic sign layouts. However, the lack of a theoretical framework for visual perception of visual capture and information conveyed by sign groups hinders the measurement of drivers’ comprehensive visual perception and the layout optimization of sign groups. To address this gap, this study introduces a calculation method for sign group information volume and a visual cognition model, establishing a comprehensive evaluation approach for sign group visual cognition. Eye movement data, collected using eye-tracking technology, were utilized to evaluate the comprehensive visual perception and optimize the layout of sign groups. The findings indicate that a low information volume fails to enhance recognition ability and alleviate the psychological burden. Conversely, excessive information may result in overlooking signs positioned on the left and top. Furthermore, drivers are unable to improve cognitive efficiency and driving safety even with self-regulation when faced with an information volume exceeding 120 bits within a 100 m span. Overall, this study demonstrates the effectiveness of the proposed method in promoting the long-term safety effect of temporary signage layouts in underground construction areas.

1. Introduction

In recent years, sustainable hydropower stations have increasingly become a public focus hotspot as rapid economic development has led to increasing environmental measures and environmental awareness [1]. Sustainable hydropower station projects often require the construction of complex underground caverns to install power generation equipment under unfavorable conditions [2]. To improve construction efficiency, multiple working surfaces are often constructed simultaneously, which leads to many branch caverns connecting to the main access tunnel of the underground caverns. Drivers have to transport sludge, materials, equipment, and personnel for extended periods in underground construction caverns [3]. Due to the intricate construction adits and the frequent interaction between construction and traffic, driving accidents have the highest rates of injury and death in underground projects [4]. These hazardous driving environments constantly threaten the safety of drivers [5]. Sustainable hydropower station construction and transportation safety are greatly challenged [6].

Signs serve as a medium for transmitting driving and safety information during the construction process of underground caverns. To ensure safe driving and protect construction personnel in the vicinity, drivers must derive valuable information from the signs present in the driving environment [7,8]. In the construction phase of underground caverns, there exist complex visual cues comprising traffic-related and construction safety-related signs. These signs are temporarily installed as the construction progresses, and their number and layout can be more intricate and chaotic than those in traffic tunnels. Unlike signs in completed traffic tunnels, signs in underground construction caverns not only convey speed limits and no-overtaking information but also provide construction reminders and direction guidance, as shown in Figure 1. Meanwhile, signs tend to appear in groups for a short period. Drivers have to extract traffic guidance and information on construction hazard avoidance from a large amount of information quickly, resulting in a higher visual load on drivers. This could hinder their ability to interpret information correctly and promptly interpret the information, ultimately leading to erroneous driving decisions [9]. Therefore, it is essential to study the optimization of sign layout during the construction process of underground caverns to minimize struck-by accidents and safeguard personnel, including drivers and workers.

The reasonable layout of signs represents an effective measure to mitigate safety issues in underground construction caverns. In the United States, the Federal Highway Administration established the Manual of Uniform Traffic Control Devices (MUTCD), which sets limits on the number of words used on traffic signs [10]. Constraints are imposed on sign groups, and it is strongly recommended that each booth containing two or three signs should display no more than one station name, with a maximum of three station names per booth. The principal consideration in sign layout design should prioritize the driver’s efficient reception and utilization of information, specifically their visual cognitive ability to process sign information [11]. Because there is a limit to the amount of information a driver can process in a short period, excessive information causes visual overload on the driver and leads to neglect of critical information [12]. Moreover, signs possess various attributes including word count, color, frame shape, and location, each of which has distinct implications for visual cognition and information transmission capabilities [13]. Consequently, sign layout must account for the disparities in conveying information between different sign attributes. Overall, the key to optimizing the layout of sign groups is to understand the driver’s visual perception, namely, the ability to capture signs and the ability to communicate the information on signs. However, most of the existing models for visual cognition focus on single factors, such as intention preference, target search, and information processing, without much consideration of the impact of visual cognitive processes on visual recognizability [14,15]. It is necessary to construct a visual cognitive model to understand the driver’s cognitive processing of signs.

Sign groups carry varying information, and different volumes of information can affect a driver’s visual recognition [16]. The information theory by Shannon uses a mathematical formula to measure the information volume, which is defined as the occurrence of an event that reduces uncertainty and is measured in bits [17]. The information volume has a linear relationship with the reaction speed of the subject [18]. Previous research on tunnel sign information volume mainly focused on traffic signs [19,20]. Meanwhile, these studies predominantly examined the information conveyed by individual signs, with limited consideration given to the influence of sign groups on drivers’ visual recognition. Sign groups formed by the combination of traffic and construction safety signs in temporary underground construction caverns are more intricate compared with completed traffic tunnels. Conventional methods for measuring information conveyed by standard traffic signs are inadequate in this context. Thus, information theory is used to establish a method for quantifying information in sign groups containing traffic and safety information and to lay the foundation for analyzing the visual capture of sign groups by drivers and the information conveyed by sign groups.

There are two main methods for assessing a driver’s ability to acquire information from signs. One approach involves inferring visual perception by measuring and evaluating driver reaction times in decision-making tasks [21]. The other approach involves exploring visual perception by collecting eye movement data in real-time using sensing technology. The method that uses drivers’ response times to analyze their ability to acquire driving information is indirect and less accurate. Vision is a sensory perception, and the most effective way to assess it is by obtaining real-time eye movement data using sensing technology [22]. In general, previous studies have focused on assessing drivers’ gaze behavior in tunnels using eye-movement data and adjusting the layout of individual traffic signs. However, there is a lack of exploration of the visual cognitive processes of sign groups. The reliability of the selected eye-movement parameters as evaluation indicators is insufficient. Therefore, in this study, a visual cognition model is constructed using visual cognition theory, which is of practical significance to exploring the evaluation index of the comprehensive visual recognition effect and revealing the optimization of sign group layout in underground construction caverns.

To fill the gap in existing studies, a method is proposed for calculating the information volume in these caverns by integrating the basic elements of traffic and safety signs. Next, a visual cognition model is established and evaluation indices for drivers’ visual recognition are determined based on this model. Finally, the comprehensive visual cognition of sign groups with different information volumes is evaluated using eye-movement data required for collecting indexes using eye-tracking, and suggestions for layout optimization are made based on the results.

This study aimed to explore the optimization of sign layouts in underground construction sites. This paper developed a calculation method for sign group information volume and a visual cognition model to determine a comprehensive evaluation method for sign group visual cognition. The comprehensive visual cognition of sign groups with different information volumes was evaluated using eye-movement data required for collecting indexes using eye-tracking, and suggestions for layout optimization are made based on the results. The following questions were addressed:

(1) Is the original study on the optimization of sign layout in traffic tunnels based on the volume of information applicable to the special environment of transportation in underground construction tunnels?

(2) Under the environment of transportation in underground construction caverns, what is the level of information conveyed by signs that is more conducive to visual cognition?

(3) Based on this study, what recommendations can be made for the signage layout in underground construction caverns to enhance the safety of construction and transportation in sustainable hydropower plants?

The remainder of this paper is organized as follows: A literature review on sign group layout, the visual cognitive model, and eye-tracking technology for sign optimization is presented in Section 2. The primary purpose is to further elaborate on the innovations of this study using a comparison with previous studies. Section 3 proposes this study’s methodology, including the experimental design, information quantification of the sign group, quantification of eye movement indices, selection of evaluation indicators, and data collection and processing. In Section 4, the results are presented and discussed. Finally, the conclusion, limitations, and recommendations for the layout of underground construction cavern sign groups are proposed in Section 5.

2. Literature Review

During construction, driving safety is always a common concern in underground caverns. Safe driving decisions require access to driving information on signs as a prerequisite. The reasonable layout of signs is one of the key measures used to alleviate safety problems in underground construction caverns. This section reviews the design and layout of the sign, the visual perception of the sign, and eye-tracking technology for sign optimization.

2.1. Design and Layout of the Sign System

Researchers have used visual cognition as an entry point to design and optimize the layout of a sign based on the visual communication effect of the sign. In research on designing and optimizing signs, scholars have concentrated on the influence of factors such as sign type, content, location, size, and constituent elements on the effectiveness of sign information communication. Notably, warning signs and speed limit signs, which are utilized to convey route characteristics, specific road maintenance conditions, and highway regulations, have the potential to influence drivers’ operational behavior [23]. Research investigating driver responses to warning sign design using public surveys has further revealed that each component of a sign, as well as its location, influences driver responses [24]. Additionally, it was discovered that the warning distance of advance guide signs for exit ramps impacts layout effectiveness [25].

Research has also shown that optimizing the amount of information conveyed by signs ensures that drivers adequately receive guidance information and make correct driving decisions [26]. Information quantification theory suggests that signs possess varying volumes of information based on their constituent elements, including color, word count, number of words, and boundary shape. These factors can affect the driver’s ability to recognize a given sign [27]. The quantification of information thresholds for road signs can be achieved with the application of the information entropy framework, which is rooted in information theory [28]. Moreover, an optimization model can be constructed for the layout of guidance signs, with the objective function aiming to maximize the information received by individuals [29]. Similarly, in the context of construction areas, optimizing traffic sign layout can be achieved by introducing a model that utilizes the objective function for maximizing the guidance information provided by the traffic signs within the construction impact zone [30].

Concurrently, various investigations were conducted from the perspective of visual perception to examine the design and layout of signs. For instance, a simulated scenario was created incorporating six commonly encountered visual guidance signs to evaluate the impact of different visual guidance signs on drivers’ visual perception [31]. Attention allocation of the subjects was quantified to examine the influence of individual signs with varying numbers and formats on visual attention [32]. The impact of layout on recognition effectiveness was assessed using distinct sign location arrangements [33]. During a simulated driving test, participants’ visual capture distance and driving performance metrics were assessed concerning signs, enabling the optimization of advanced guidance distances [34]. After collecting visual and operational behavior parameters of drivers using driving simulation tests, the combined layout of warning signs and deceleration measures at tunnel entrances was analyzed [35]. Additionally, a virtual reality experiment was conducted in a constructed building information modeling (BIM) model scenario to compare human visual perception before and after implementing the optimized sign layout, thereby evaluating its effectiveness [36].

Most past studies have focused on optimizing the layout of individual signs on completed roads, tunnels, shopping centers, and underground parking lots. These optimal layouts are studied based on the observer’s visual perception and information processing of the sign. In underground construction caverns, drivers are often required to drive long distances in dark environments while quickly extracting critical information from complex traffic and safety signs. Whether the results of previous studies are also applicable to the sign group layout of underground caverns in construction environments remains to be further verified.

2.2. Human Visual Perception of a Sign

The ultimate purpose of sign layout is to guide drivers to travel safely. The visual capture of the sign and the reception and processing of the information it conveys become key. Researchers have found that these two points have a close connection with a driver’s visual cognitive process. Common cognitive models point to attention to the allocation of drivers’ visual attention [37], information processing [38], and cognitive load [39] as key influences on cognitive ability. Visual perception is the primary way in which a person perceives objective information about the external environment.

Improper allocation of drivers’ visual attention to signs and road conditions is widely recognized as a significant contributor to traffic accidents [40]. The demanding nature of driving, coupled with the presence of complex road conditions and environments featuring mixed signs with advanced guidance information often hinders drivers’ ability to effectively observe road conditions and allocate their gaze accordingly [41]. The prediction of drivers’ attention allocation has been achieved with the utilization of HSMM and RNN models [42]. To characterize driver behavior within a visual–cognitive distraction space, a combination of regression models with elastic net regularization and binary classifiers has been used [43]. Moreover, an attention-guided model, constructed using visual features acquired with convolutional neural networks (CNNs), has been utilized to classify objects in a goal search experiment, highlighting the role of sign-guided attention [44].

Meanwhile, it is vital to ensure that drivers can effectively process the information conveyed by signs. The development of attention-shifting and prediction models can enhance our understanding of sign recognition rates and times. For instance, the observation of a driver’s gaze shift pattern has been accomplished with the application of hidden Markov models [45]. Another hidden Markov model-based cognitive model, focusing on covert visual attention, has been used to investigate the switching paradigm between two covert attention states (local and global attention) in complex scenes, thereby facilitating effective sign recognition [46]. Additionally, a computational model exploring visual perception and intermittent attention has been constructed to examine a driver’s capacity to process information based on gaze duration [47].

Simultaneously, the reception level of information conveyed by signs is closely associated with a driver’s visual cognitive load [48]. Insufficient cognitive load can result in diminished alertness, while cognitive overload can impair information processing, leading to decision-making or operational errors and an increased risk of traffic accidents [40]. Thus, assessing cognitive load can unveil the impact of sign set layout on driver cognitive load and offer guidance for optimizing such layouts [49]. Several methods can measure drivers’ cognitive load, including subjective evaluation, driving task performance, and physiological parameter evaluation [50]. Subjective evaluation methods are susceptible to individual variances and cannot provide real-time assessment of mental load. On the other hand, implementing the driving task performance method in actual traffic environments is challenging. The physiological parameter evaluation method indirectly assesses drivers’ psychological load by analyzing changes in their EEG, ECG, eye movements, and other indicators [14]. As psychological load fluctuations are accompanied by corresponding changes in physiological and psychological indicators, the physiological parameter evaluation method allows for real-time, objective, and highly sensitive evaluation of the psychological load [51,52]. Machine learning models and deep learning architectures have been used to classify and predict drivers’ cognitive load based on eye movement signals [53]. Moreover, an information transmission model for signs has been developed from the perspective of cognitive load to examine the relationship between driver recognition time and different traffic signs [54].

Current research on sign layout has primarily focused on a driver’s visual cognitive model at the macro level while neglecting to explore the comprehensive visual cognitive process encompassing attention to road conditions, visual capture of signs, and processing of sign information during the driving process. There is a lack of research on establishing evaluation systems to measure drivers’ comprehensive visual perception and optimizing the layout of complex sign groups.

2.3. Eye-Tracking Technology Application for Sign Layout Optimization

The layout of a sign needs to consider the efficiency of information communication and a driver’s ability to visually capture the sign. Both of these need to be reflected in a driver’s visual cognitive behavior. Scholars have found that eye movement measures are often used as a statistical factor related to potential concepts of visual cognition [55]. Visual cognition can be reflected by changes in eye movement behaviors such as blink rate, pupil diameter, duration, and extent of visual field [56]. Eye-movement techniques are the most direct way to obtain these eye-movement data [57].

Regarding information communication, scholarly research has predominantly focused on using eye-movement techniques to capture various eye-movement parameters. This approach facilitates an analysis of how the fundamental elements of a sign influence information transfer and subsequently explores the optimization of sign layouts. For instance, eye movement tests have been utilized to analyze drivers’ comprehension and subtask performance when exposed to traffic safety messages on dynamic signs [58]. The objective was to investigate the impact of sign content and word count restrictions on the layout. Simulation experiments were conducted using a driving simulator and an eye-tracking device to assess the effectiveness of utilizing two distinct shapes of deceleration signs for conveying sign messages [59]. Furthermore, the effectiveness of the scheme for exit warning signs in mountain tunnels was evaluated using the aforementioned simulation experiments and eye-tracking technique. The aim was to optimize the positioning of advanced guidance signs within mountainous road tunnels [60]. Using eye-tracking techniques, parameters related to drivers’ eye-tracking characteristics were extracted, allowing for an analysis of the effects of drivers’ age and the information layout on visual perception and comprehension of variable message symbols [61].

Regarding visual capture ability, research has predominantly focused on using eye-tracking techniques to extract eye movement parameters for analyzing the impact of signs on drivers’ visual attention. Consequently, researchers have sought to optimize sign layouts. For instance, eye-tracking technology was used to gather eye movement data during a simulated driving test, enabling an investigation into drivers’ visual recognition time and behavior with factor analysis and gray correlation analysis [62]. While manipulating the positions of various speed limit signs, researchers observed drivers’ eye movement behavior and driving speed, thereby uncovering the influence of sign placement on drivers’ visual attention across different scenarios [63]. Furthermore, by combining eye movement data with Markov models, the visual search paths of drivers, both familiar and unfamiliar, were analyzed specifically concerning guide signs positioned at road tunnel exits. That analysis contributed to the development of a proposed method for sign layout [12]. Additionally, the distribution of eye gaze time between the road and signs was quantified to evaluate the effect of signs on visual attention and driving safety [64]. To assess the efficacy of visually guided sign layouts, eye movement data on scanning patterns, the accuracy of speed perception, and reaction time were collected as indicators of driver behavior [65].

In summary, scholars have studied the basic sign layout of traffic tunnels and construction areas using eye-tracking techniques. However, few scholars have considered the effects of complex sign groups on driver visual perception in underground construction caverns. The application of eye-tracking technology offers the possibility to explore the influence of the above-mentioned factors on drivers’ visual perception.

3. Materials and Methods

Typical visual cognition models, which analyze sign visual cognition from individual points such as visual attention preference, visual search, and information processing, ignore the influence of cognitive processes on recognition effects. It is necessary to construct a visual cognition model based on the visual cognition process to better explore the visual cognition effect of sign arrangement. The development and application of eye-tracking technology have enabled researchers to obtain eye movement data to analyze the visual cognitive process. In addition, eye movement can intuitively reflect the driver’s visual perception behavior [66]. The best way to analyze visual behavior is to collect eye movement data in real-time using sensing technology. In recent studies, the use of eye-tracking technology has become a familiar and effective method for collecting eye movement data [67]. Therefore, using eye-tracking experiments, the visual cognitive process of drivers in underground construction cavern vehicles can be well analyzed to obtain eye movement data during the process of sign recognition.

3.1. Experimental Design

Driving simulation is a tool that provides an alternative to on-road testing in a safe, reliable, and efficient manner. Due to the danger of driving in construction sites, an actual vehicle experiment would have presented a significant safety risk. In this paper, environmental video data on the access tunnel in an underground cavern of a hydropower station were collected and imported into a driving simulator. Next, a simulated driving experiment was performed, and participants’ eye movement data were collected using eye-tracking in real-time. Before the beginning of the experiment, participants were asked to observe visual cues during simulated driving. After the experiment, a semistructured open-ended interview was conducted to determine the drivers’ attention deviation for different signs. The participants were required to simulate driving at a speed of approximately 40 km/h. In this section, the experimental design of this study, including the experimental scenarios, apparatus, participants, and procedure, is described in detail.

3.1.1. Experimental Scenarios

The access tunnel in the underground construction cavern of Jinping Hydropower Station was selected as the research object in this study. This project is located at the bend in the Jinping River on a tributary in the Yalongjiang River in Southwest China. The access tunnel was set up to assist with the construction of the underground cavern by connecting the construction adit and providing a path for discarding slag and transporting muck, materials, equipment, personnel, etc. The total length of the access tunnel is 17.5 km, and the access tunnel speed limit is 40 km/h. In this study, environmental data from the construction phase of an underground cavern at this hydropower plant were collected and imported into a driving simulator. Simulated driving experiments were conducted in a driving simulation laboratory, and eye movement data from the participants were collected in real-time using eye-tracking technology. To simulate the dark environment in underground caves, the experiment was conducted at night with all light sources turned off.

The driving process in the main access tunnel of the underground construction cavern was selected as the research object, focusing on the visual cognitive characteristics of drivers in the underground construction area. The data on the driving process was divided into a total of 175 segments, with each segment spanning 100 m, as shown in Figure 2a. This paper checked the validity of data segments, eliminated data segments without a sign, and finally, extracted valid data for 68 scenarios, as shown in Figure 2b. All signs present in each scene are considered a sign group. In the same experimental environment of an underground construction cavern, eye movement data in each valid scene was analyzed. The variation in drivers’ visual perception of signs in the access tunnel of the underground cavern was discussed.

3.1.2. Experimental Apparatus

The experimental instruments included a laptop computer, a driving simulator for the construction vehicle model DH-QCD-01, and a Tobii Pro Glasses 2 eye-tracking device, as shown in Figure 3. The driving simulator included a central unit, monitor, steering wheel, throttle, brake, clutch, and shift lever. The accuracies of the eye-tracking device for the visual angle and head displacement were 0.5 degrees and 1 mm, respectively [68].

3.1.3. Participants

In this study, a demographic questionnaire was used to collect information on the participants’ gender, age, driving age, and tunnel driving experience. Subjects were required to have a corrected visual acuity of 5.0. Additionally, none of the participants in this sample had previous driving experience in the experimental tunnel. Fourteen subjects who met the requirements of the experiment were selected. All subjects were required to have a corrected visual acuity of no less than 1.0. At the same time, all subjects had driving experience in underground cavern engineering but were not familiar with the driving simulation scene. On the day of the experiment, the subjects were in a good mental state and did not drink stimulating drinks such as tea or coffee to reduce the impact of other factors on the subjects.

After collecting eye-movement data from the first five participants, the sample size was reevaluated using a one-way ANOVA (F-test) with PASS 15.0.1 software to ensure the accuracy and reliability of the sample size. α was set at 0.05 (power of 80%). In the scene where the sign group appeared, two subjects with the most null data were excluded. Finally, eye movement data were collected in the underground cavern environment based on 12 subjects during the driving simulation experiment, as shown in

Table 1. Descriptive demographic statistics of participants.

Subject	Age	Sex	Driving Experience in Underground Caverns	Driving Experience in This Test Scene	Driving Mileage/km	Types of Vehicles Frequently Driven	Percentage of Null Data
1	52	M	Yes	No	≥50,000	Medium trucks	11.02%
2	41	M	Yes	No	≥50,000	Dump trucks	8.11%
3	38	F	Yes	No	<5000	Dump trucks	7.92%
4	22	M	Yes	No	5000–50,000	Light trucks	32.11%
5	36	F	Yes	No	5000–50,000	Light trucks	17.21%
6	53	M	Yes	No	≥50,000	Dump trucks	6.49%
7	27	F	Yes	No	5000–50,000	Light trucks	8.44%
8	23	F	Yes	No	<5000	Light trucks	12.79%
9	51	M	Yes	No	<5000	Medium trucks	16.25%
10	26	M	Yes	No	<5000	Light trucks	11.65%
11	58	F	Yes	No	≥50,000	Dump trucks	5.33%
12	46	F	Yes	No	5000–50,000	Light trucks	2.18%
13	29	M	Yes	No	≥50,000	Dump trucks	14.21%
14	21	F	Yes	No	5000–50,000	Dump trucks	27.28%
Mean	37.36
SD	12.47

Bold indicates the two subjects with the highest percentage of null data. M means male; F means female.

.

Eye movement data from the final 12 participants, with ages ranging from 23 to 58 years (mean = 40.00; SD = 11.51) were used for analysis. The experiment recruited participants with driver’s licenses, including 6 novice and 6 skilled drivers. The types of vehicles that the participants most frequently drove were light trucks (5 subjects), medium trucks (2 subjects), and dump trucks (5 subjects). Participants were required to have a corrected visual acuity of at least 1.0 and have driving experience in underground caverns. They were unfamiliar with the driving simulation scene. Participants signed a statement indicating their voluntary participation in the experiment and received a small gift as acknowledgment.

3.1.4. Experiment Procedure

This study commenced with a demographic survey questionnaire to determine the suitability of potential subjects based on their gender, age, driving age, and tunnel driving experience. Subsequently, a preliminary experiment was carried out using eye-movement equipment in a similar tunnel environment scenario. The eye-movement parameters from two randomly selected subjects were collected and analyzed to assess the feasibility of the experiment. Additionally, areas of interest for gaze were pre-identified to facilitate the identification of complex visual cues during the data analysis stage. Furthermore, the influence of the six basic elements of signs on the message conveyed in the underground cave room was studied using a questionnaire.

Before the experiment, participants were briefed on the proper use of the equipment and experimental procedures. Each participant wore Tobii Pro Glasses 2 while seated on a simulator. A one-point eye calibration was conducted to ensure precise gaze tracking. Following the calibration, the lights in the driving simulation lab were turned off to replicate the dim environment of an underground cavern, and the simulation experiment commenced. During the test phase, the investigator did not provide instructions or take any action that could interfere with the test. After the simulated driving test, the recording was stopped, and the validity of the driver’s test was confirmed using playback before saving the data.

Then, participants took a five-minute break, and a post hoc semi-structured interview was conducted to explore the reasons for each driver’s attention deviation from various signs during the experiment. The interview outline consisted of questions such as: Did you focus more on traffic or safety signs? Did you receive all the cues about driving? What factors influenced you to look for all the signs in the sign cluster? What factors influenced your reading of the sign information? Please give examples.

Finally, using data analysis and an evaluation of the effectiveness of visual cognition, the rationality of the layout of the sign group was analyzed, and some layout suggestions were proposed. The framework of research ideas in this paper is shown in Figure 4.

3.2. Quantification of Information Conveyed by Sign Groups

3.2.1. Calculation of Information Volume for Basic Elements of Signs

As a cognitive carrier in the driving environment, signs are transmitted to drivers using symbols, words, colors, and other basic elements. The driver’s acquisition of information is a process from ignorance to cognition. Based on the information theory of Shannon [17], the information volume is related to the number of states and the probability of state occurrence. The average volume of information per symbol is related to the probability of occurrence for each essential element of the sign. For a set of events X, the sign information can be expressed using Equation (1):

$$H(X)=-{\displaystyle \sum _{i=1}^{m}P}\left(X_i\right){\log }_{2}P\left(X_i\right)=a_{1}H(X_1)+a_{2}H(X_2)+\cdots +a_{i}H(X_i)$$

(1)

where H(X) is the information content of traffic engineering facilities; m indicates the number of events in the set; X_i indicates the ith event; and a_i indicates the number of identifiers with corresponding information.

There are many marking elements and discrete data in the construction stage of underground caverns. To obtain the basic elements of common signs in underground tunnels and further quantify the information conveyed by these elements, several Chinese government documents were referenced, including the Table of General Standard Chinese Characters [69], Road traffic signs and markings: Part 1: General [70], Road traffic signs and markings: Part 2: Road traffic signs [71], Road traffic signs and markings: Part 4: Work zone [72], and Safety signs and guidelines for the use [73], as shown in

Table 2. Basic information volume of common signs in underground caverns.

No.	Type	Total Number in the Standard/Total Number in the Cavern					Total (without Repeats)	Basic Information Volume/Bits
		Document 1 a	Document 2 b	Document 3 c	Document 4 d	Document 5 e
1	Chinese character	3500/3500	-/-	-/-	-/-	-/-	3500	11.773
2	Arabic numeral	-/-	10/10	10/10	10/10	10/10	10	3.322
3	Geometry of the border	-/-	-/-	6/6	8/8	3/3	6	2.585
4	Color	-/-	-/-	10/10	10/10	6/6	10	3.322
5	Pointing symbol	-/-	7/7	-/-	-/-	-/-	7	2.807
6	Graphic or symbol	-/-	75/55	289/134	11/10	103/103	247	7.948

^a Table of General Standard Chinese Characters; ^b Road traffic signs and markings: Part 1: General; ^c Road traffic signs and markings: Part 2: Road traffic signs; ^d Road traffic signs and markings: Part 4: Work zone; ^e Safety signs and guidelines for the use.

.

In the process of information transmission, the importance of each type of element is different. To obtain the effective information transmitted by each type of element, the method suitable for measuring the effective information on temporary signs in underground caverns was determined. A questionnaire survey was conducted on the importance of the influence of six underground cavern sign elements on information transmission to the subjects. The analytic hierarchy process (AHP) was used to establish the factor set of comprehensive evaluation and set up underground caverns [74]. The weights of each element were determined to be: Chinese characters (0.215), Arabic numeral (0.118), geometry of the border (0.152), color (0.594), pointing symbol (0.197), and graphic or symbol (0.139). The judgment matrix satisfies the consistency test (CR = 0.095 < 1), as shown in

Table 3. Influence on the guiding role of signs.

	Chinese Character	Arabic Numeral	Geometry of the Border	Color	Pointing Symbol	Graphic or Symbol	ωi
Chinese character	1	3	2	1	1	1	0.215
Arabic numeral	1/3	1	1	1	1/2	1	0.118
Geometry of the border	1/2	1	1	1/2	1/2	3	0.152
Color	1	1	2	1	1	1	0.179
Pointing symbol	1	2	2	1	1	1	0.197
Graphic or symbol	1	1	1/3	1	1	1	0.139
CR	0.095

.

After assigning weights (ω_i) to the importance of the six types of sign elements in the underground cavern, the effective information volume of each element in the temporary sign in the underground cavern was calculated using Equation (2):

$$H\left(S\right)={\displaystyle \sum _{i=1}^m{a}_{i}H\left(X_i\right)}{\omega }_i$$

(2)

The effective unit information in each element of the temporary sign in the underground cavern is displayed in

Table 4. Valid information volume of common signs in underground caverns after weighting using AHP.

Type	Chinese Character	Arabic Numeral	Geometry of the Border	Color	Pointing Symbol	Graphic or Symbol
N	3500	10	6	10	7	247
Basic information volume /bits	11.773	3.322	2.585	3.322	2.807	7.948
Wi	0.215	0.118	0.152	0.179	0.197	0.139
Effective information volume /bits	2.532	0.390	0.394	0.594	0.553	1.107

.

3.2.2. Levels of Information Volume for Sign Groups

The sign group elements in the construction stage of underground caverns for five typical hydropower projects in China were investigated and counted. The results show that during the construction phase of the underground cavern, the information volume of a sign group within 100 m ranges from 1.74 to 215.01 bits, with 95.59% concentrated from 1.74 bits to 173.28 bits. In the test scenario, the information volume in the sign group within 100 m ranges from 1.74 bits to 203.74 bits. To better observe the influence of different information on drivers’ visual recognition, the sampled groups of signs were divided into 10 levels (L1 ≤ 20 bits, 20 bits < L2 ≤ 40 bits, 40 bits < L3 ≤ 60 bits, 60 bits < L4 ≤ 80 bits, 80 bits < L5 ≤ 100 bits, 100 bits < L6 ≤120 bits, 120 bits < L7 ≤ 140 bits, 140 bits < L8 ≤160 bits, 160 bits < L9 ≤180 bits, and L10 > 180 bits) [75]. Within each level of information volume, one instance was extracted, as shown in Figure 5. Meanwhile, the signs contained in each instance were also collated, as shown in Figure 6. The content in parentheses is the sequence number of the selected instance in the test material.

3.3. Evaluation of Visual Cognition

3.3.1. Cognitive Model

To better understand the visual cognitive process of transport vehicle drivers in underground construction caverns, it is necessary to construct and optimize a visual cognitive process model for analysis. Cognitive process models that are familiar to scholars include the Rasmussen cognitive ladder model [37], the Jeelani attention preference model [76], the Drury visual search model [38], the Tipper information processing model [77], the Debue cognitive load model [39], etc. The Rasmussen cognitive ladder model divides the cognitive process into eight stages: activation, observation, recognition, interpretation, evaluation, task definition, protocol formation, and execution response. The Jeelani attention preference model points out the impact of visual attention preference on the subject’s visual cognition during the activation and observation stages. The Debue cognitive load model considers the influence of psychological factors on cognitive subjects and adds the information load received by subjects to the visual cognitive model. The Drury visual search model assumes that visual capture of task targets is the key to influencing cognition. If the searcher does not allocate gaze time to the target, they will not be able to search for these areas.

In underground construction caverns, construction vehicle drivers scan their surroundings, from which they observe the road conditions and extract advance guidance information from signs to make decisions for safe driving. Therefore, perceiving signs can be essentially considered as a multi-objective visual search process, where signs are the main target of the search. However, the visual system has a limited capacity to process information about the external environment, and when the information exceeds the cognitive load, the driver cannot perceive and process all objects or targets in the external world at the same time. In the process of visual cognition, it is crucial to extract valid information and allocate attention appropriately. A construction vehicle driver may need to capture and recognize sign information in the environment completely and quickly while also allocating a certain volume of attention to the road ahead. To better understand the visual cognitive process of sign recognition and analyze the factors affecting recognition, it is necessary to analyze the cognitive process of sign recognition and establish a model for the visual cognitive process of sign recognition.

Based on the above principles of visual cognition for safe driving and the actual situation of the cave transportation environment, drivers need to first have a concern for road conditions and the environment, then capture signs guiding safe driving from the environment, and then receive and process the information conveyed by the signs within the visual load range, ultimately making decisions. Therefore, this study extracted four main cognitive links, including the perception of the environment, visual capture of signs, processing of information, and decision-making, to construct a visual cognition model for underground construction caverns, as shown in Figure 7. According to the characteristics of different cognitive links and the influence of mental load on visual cognition, evaluation indexes were selected to measure the performance of drivers’ visual recognition of signs.

3.3.2. Selection of Evaluation Indicators

According to the visual cognitive model, the aim of this study is to understand the driver’s visual cognitive process when driving an underground construction cavern transport vehicle, including the environment perception stage, visual capture of signs stage, processing of sign information stage, and decision-making stage.

Since eye movements and attentive cognition are linked, it is possible to detect drivers’ cognitive states in situ using eye trackers. Once these cognitive states are understood, the information volume in the sign group can be optimized to adapt to the driver’s current cognitive abilities (e.g., cognitive load). Fixation is the behavior of visual activity in which the central fovea of the retina in the eye is aligned with a target object for more than 100 ms [78]. Therefore, the evaluation indicators of the visual recognition process can be characterized by quantifying eye movement parameters.

(1) Environment Perception Stage

During transportation tasks, drivers need to pay attention not only to the guidance information on signs but also spend a certain amount of energy focusing attention on the road conditions and environment using gaze behavior. ADN is the average duration that participants focus on non-areas of interest (non-AOI) in different scenarios. The concern for road conditions (CRC) can be described using the ADN, which reflects the driver’s attention to the surrounding road conditions and environment. The lower the ADN, the lower the CRC, and the greater the driving risk. The function is defined in Equation (3).

$$AD{N}_{mj}=\frac{{\displaystyle \sum _{i=1}^{n}AN{D}_{mji}}}{n}=\frac{1}{n}{\displaystyle \sum _{i=1}^n{\displaystyle \sum _{j=1}^k\left(E\left(f_{mji}\right)-S\left(f_{mji}\right)\right)}}$$

(3)

where ADN_mji is the total duration of all gaze behavior of the ith subject on the jth non-AOI in the mth scene; E(f_mi) and S(f_mi) represent the end time and start time, respectively, of the single gaze behavior of the ith subject on the jth non-AOI in the mth scene; and k is the gaze behavior number of the ith subject on the jth AOI in the mth scene.

(2) Visual capture of signs stage

In the visual capture of signs stage, the impact of long-term driving in an underground construction environment on the driver’s attention heatmap and the impact of the number of signs on the recognition rate have a significant impact on the capture of signs. The spatial preference for visual attention (SPVA) can be characterized using the heat map, and the recognition rate of the sign group (RRSG) can be calculated. Using the recognition rate of AOIs, the driver’s visual perception of AOIs can be judged [76]. The higher the NCA, the higher the RRSG and the stronger the information capture ability of drivers. The function is defined in Equation (4).

$$RRS{G}_m=\frac{{\displaystyle \sum _{i=1}^n\frac{NC{A}_{{}_{mi}}}{N{A}_{{}_{mi}}}}}{n}$$

(4)

where NCA_mi is the number of captured AOIs (signs) by the ith subject in the mth scene; NA_mi is the total AOIs for the ith subject in the mth scene; n is the number of subjects.

(3) Processing of sign information stage

In the processing of sign information stage, the processing and cognitive ability of information (PCAI) are the key issues [79]. The long average duration of AOI (ADA) of gaze behavior indicates that this area contains complex information, and recognition is difficult [61]. Hence, PCAI can be described using the ADA, which reflects the time drivers spend processing the information. The higher the ADA, the lower the PCAI. The function is defined in Equations (5) and (6).

$$PCAI=\frac{1}{ADA}$$

(5)

$$AD{A}_{mj}=\frac{{\displaystyle \sum _{i=1}^{n}AD{A}_{mji}}}{n}=\frac{1}{n}{\displaystyle \sum _{i=1}^n{\displaystyle \sum _{j=1}^k\left(E\left(f_{mji}\right)-S\left(f_{mji}\right)\right)}}$$

(6)

where ADA_mji is the total duration of all gaze behaviors of the ith subject on the jth AOI in the mth scene; E(f_mi) and S(f_mi) represent the end time and start time, respectively, of the single gaze behavior of the ith subject on the jth AOI in the mth scene; and k is the gaze behavior number of the ith subject on the jth AOI in the mth scene.

(4) Decision-making stage

In the decision-making stage, the driver’s decision-making could be affected by the above three cognitive processes and the visual cognitive load. In particular, the impact of cognitive load on decision-making. There is strong evidence that a longer average fixation duration is related to a higher cognitive load [80], which indicates that more attentional resources are required [39]. The pupil diameter is a sensitive indicator of cognitive load, and pupil dilation indicates a large cognitive load [81]. Therefore, this study selected changes in the mean diameter of the pupil (MDP) to characterize the driver’s mental workload. The larger MDP, the greater the ML.

$$MD{P}_{mi}=\frac{{\displaystyle \sum _{i=1}^k{B}_{mi}^1+\cdots +B_{mi}^k}}{k}$$

(7)

where B is the diameter of the pupil; and k is the number of fixation points of the ith subject in the mth scene.

Thus, this paper selected some indicators, namely, the fixation heatmap, the average duration of AOIs (ADA), the number of AOIs (NA), the number of AOIs captured (NCA), the average duration of non-AOI (ADN), and the mean diameter of the pupil (MDP). These metrics were used to construct three evaluation indicators to evaluate the driver’s visual recognition and mental load based on the different information volumes of the sign group in underground construction caverns, as shown in

Table 5. Selection of evaluation indicators based on eye movement data.

Evaluation Indicators	Eye Movement Indicators	Description	Interpretation of Indicators
Spatial preference for visual attention (SPVA) [82]	-	The fixation heatmap provides a more visual representation of the spatial allocation of subjects’ attention to the sample content.	The red area indicates the highest level of attention of the subjects, the yellow area is the second highest, and the green area is the lowest.
Concern for road conditions (CRC) [83]	Average duration of non-AOI (ADN)	The CRC can be described using the ADN, which reflects a driver’s attention to the surrounding road conditions and environment.	The lower the ADN, the lower the CRC and the greater the driving risk.
Recognition rate of the sign group (RRSG) [84]	Number of AOIs (NA); Number of captured AOIs (NCA)	$RRSG=\frac{NCA}{NA}$, where RRSG represents the drivers’ ability to capture information.	The higher the NCA, the higher the RRSG and the stronger the information captured ability of drivers.
Processing and cognitive ability of information (PCAI) [79]	Average duration of AOIs (ADA)	$PCAI=\frac{1}{ADA}$, where PCAI can be described using the ADA, which reflects the time drivers spend processes the information.	The higher the ADA, the lower the PCAI.
Mental load (ML) [85]	Mean diameter of the pupil (MDP)	The ML can be described using the MDP, which is a sensitive indicator of mental load.	The larger MDP, the greater the ML.

.

3.3.3. Evaluation Method for the Comprehensive Effect of Visual Cognition

To thoroughly examine the patterns of change in drivers’ visual cognitive effects, it is advantageous to consolidate multiple evaluation indices into a unified comprehensive index.

Firstly, we ensured that all evaluation metrics are isotropic. The negative eye movement indicator ADA was transformed into a positive indicator ADA⁻¹ to quantify PCAI (perceptual cognitive attention index). Subsequently, we standardized the data for each indicator obtained during the experiment to improve comparability among data indicators. The suitability of principal component analysis (PCA) for this study was assessed using the Kaiser–Meyer–Olkin (KMO) measure and Bartlett’s sphericity test. A PCA was conducted, and the principal components with eigenvalues exceeding 1.000 and a contribution rate of over 60% were selected to effectively represent the original data.

The expressions of each principal component were derived from the component score coefficient matrix as follows:

$$\{\begin{array}{c}{F}_1=a_{11}{Z}_1+a_{12}{Z}_2+\dots +a_{1j}{Z}_j\\ ⋮\\ F_n=a_{n1}{Z}_1+a_{n2}{Z}_2+\dots +a_{nj}{Z}_j\end{array}$$

(8)

where n is the number of all principal components, that is, the number of components with eigenvalues exceeding 1.000 and contribution rates exceeding 60%; j is the number of evaluation indicators; a_nj is the score coefficient of the jth indicator in the nth component; and Z_j is the value of the jth indicator after standardization conversion.

Using the percentage of variance contribution and total cumulative contribution corresponding to each principal component as weights, the visual perception intensity (F) of drivers in underground construction cavities for groups of signs with different information content was quantified as follows:

$$F=b_1{F}_1+b_2{F}_2+\cdots +b_n{F}_n$$

(9)

where b_n represents the ratio of variance contribution percentage and total cumulative contribution percentage corresponding to the mth principal component, namely, weight.

Finally, we completed a comprehensive evaluation of the drivers’ visual cognitive effects in underground cavern construction by quantifying the value of F under different information volumes.

4. Results and Discussion

4.1. Data Analysis

4.1.1. Data Collection and Preprocessing

The experiment was conducted with a simulated driver, and eye movement data were collected simultaneously with the wearable eye-tracking device Tobii Pro Glass 2. This test checked the validity of data segments, eliminated data segments without a sign, and finally extracted valid data for 68 scenarios. The eye movement data were then exported and initially processed using ErgoLAB 3.0, a human factor engineering analysis software.

Then, the eye movement parameters were defined using the ErgoLAB 3.0 software. A double-threshold eye movement detection method was used to mark rapid eye movement (REM) events: the onset of a REM was defined as the moment gazes velocity exceeded a 50°/s threshold value; multiple eye movements within 0.05 s were merged into one movement, and the minimum duration of an eye movement was set at 0.1 s [86]. Pupil data blinks were treated with linear interpolation, and the resulting pupil traces were low-pass filtered and smoothed following the conventions outlined in [87].

Before the experimental data analysis, the experimental data were screened to exclude missing and significantly biased data. Anomaly detection was conducted according to the three-sigma principle (3σ), and the abnormal records were removed.

4.1.2. Delineation of AOIs

To further explain the effect of different levels of information volume on drivers’ vision, this study attempted to quantify the visual recognizability of the sign group using the delineation of areas of interest. Previous research paradigms have focused on the mechanical delineation of AOIs based on the driver’s visual location of planar coordinates [78,88]. Few of these studies used visual cues as a condition for delineating AOIs, and the effects of different signs on driver attention have not been adequately considered. In this study, 68 scenes with different information volumes of sign groups (AOIs) were extracted from the access tunnel of the underground construction cavern. Then, each sign was used as the basis for dividing AOIs in scenes, as shown in Figure 8.

4.1.3. Production of Mapping Materials

To better illustrate the SPVA, it is necessary to visualize the fixation points. An attentional heat map provides a more visual representation of the spatial distribution of subjects’ attention to the sample content. Generating a heat map of gaze points requires the creation of a mapping image. In this study, each scene was synthesized separately as a static composite map of mapping material, as shown in Figure 9.

4.1.4. Data Processing and Statistical Analysis

After removing abnormal values, the eye movement data were further processed and exported using ErgoLAB 3.0 software. These data included the type of eye state, the number and duration of fixation, and the MDP. The ADA, NCA, and ADN based on the AOI and non-AOI were also included. Simultaneously, the RRSG was summarized by calculating the ratio of NCA to NA. Because the NA encountered by each participant in the experimental scenario is fixed, a statistical analysis of NA is not considered here.

These eye-movement indicators were explored using the Shapiro–Wilk normality test and the homogeneity of variance test using SPSS Statistics 26. The statistics and homogeneity test results are shown in

Table 6. Statistics, homogeneity, and one-way ANOVA test results.

Index	Mean	SD	N	Homogeneity Based on the Mean				One-Way ANOVA
				LS a	df1	df2	Sig.	F	p
NCA	4.69	4.15	816.00	2.72	9.00	806.00	0.05	17.18	0.00 **
ADN	0.41	0.24	816.00	8.93	9.00	806.00	0.06	15.80	0.00 **
ADA	0.51	0.40	816.00	7.68	9.00	806.00	0.07	5.15	0.00 **
RRSG	0.94	0.13	816.00	25.99	9.00	806.00	0.26	29.13	0.00 **
MDP	3.93	0.71	816.00	0.94	9.00	806.00	0.09	3.91	0.00 **

^a LS is the Levene statistic. ** The mean difference is significant at the 0.01 level.

. The results indicated that normal distribution (p > 0.05) and variance homogeneity (p > 0.05) were confirmed for the values of NCA, MDP, ADA, ADN, and RRSG.

Then, a one-way ANOVA was used to explore the differences between each eye movement indicator for different levels of information volumes, as shown in Table 6. The results suggested that the NCA (F = 17.18, p = 0.00 < 0.01), ADN (F = 15.80, p = 0.00 < 0.01), ADA (F = 5.15, p = 0.00 < 0.01), RRSG (F = 29.13, p = 0.00 < 0.01), and MDP (F = 3.91, p = 0.00 < 0.01) indicators were significantly different between groups.

4.2. Analysis of Evaluation Indicators of Visual Cognition

4.2.1. Spatial Preference for the Driver’s Visual Attention

The experiment generated a large volume of abstract, numerical-based eye movement data. However, these data were difficult to understand intuitively. Therefore, a fixation heatmap was used to convert the gaze data into images displayed on a screen, which intuitively expressed the distribution characteristics of gaze points. To intuitively analyze the gaze heatmaps, the environmental information in every scene was processed into a static composite graph as a mapping graph. On these mapping graphs, the eye movement fixation points were superimposed and calculated. Then, the heatmap of fixation points was generated, as shown in Figure 10.

As shown in Figure 10a–f, the driver’s gaze points were mainly distributed in front of the field of view. Meanwhile, visual cues in some locations, such as standing water on the ground, signs above the cave roof, and signs on the left and right sides, also attracted attention. This indicates that the driver captures information from the sign group while maintaining a certain level of attention to the road condition when the information volume in the sign group is less than L6. As shown in Figure 10g–j, the fixation on signs on the right side increased, while the fixation on signs on the left side decreased. Meanwhile, the fixation on signs above the cave roof and road conditions also decreased. This indicates that as the information volume increases, drivers devote their main attention to the identification of signs. There is a preference for this attention, which is mainly concentrated on the right side. Meanwhile, attention to the environment and road conditions gradually decreases, and the driving risk gradually increases.

The results differ significantly from existing studies on visual cognition during tunnel driving [89]. In the existing studies, drivers’ attentional preference tends to be more to the right and does not correlate significantly with the volume of information on the sign. The reason for this discrepancy is thought to be that ordinary transportation tunnels are monotonous, and without the dynamic influence of construction factors, drivers’ visual attention patterns are gel fixed. The transportation environment of underground construction tunnels is complex and variable, and drivers need to be constantly alert to the information conveyed by signs. When the amount of information conveyed by the signs exceeds the load, the driver’s attention preference gradually converges to the habit. That is, they ignore the left and top signs.

4.2.2. Concern for Road Conditions

During driving tasks, to ensure safety, a driver’s gaze includes both attention to specific objects in the environment (signs and road conditions) and attention to visual distractions. ADN reflects the driver’s attention to the surrounding road conditions and environment. The lower the ADN, the lower the attention to road conditions and the greater the driving risk. Multiple comparisons of ADNs at different levels of information volume were performed using one-way ANOVA.

It was found that the p-values of ADNs between L1 and L3 (p = 0.09), L1 and L6 (p = 0.91), L1 and L7 (p = 0.57), L1 and L8 (p = 0.23), L2 and L6 (p = 0.45), L2 and L7 (p = 0.14), L2 and L6 (p = 0.45), L3 and L4 (p = 0.36), L3 and L5 (p = 0.65), L3 and L6 (p = 0.45), L3 and L7 (p = 0.55), L3 and L8 (p = 0.95), L4 and L5 (p = 0.67), L4 and L6 (p = 0.19), L4 and L7 (p = 0.17), L4 and L8 (p = 0.44), L4 and L10 (p = 0.10), L5 and L6 (p = 0.30), L5 and L7 (p = 0.32), L5 and L8 (p = 0.67), L5 and L10 (p = 0.07), L6 and L7 (p = 0.80), L6 and L8 (p = 0.52), L7 and L8 (p = 0.64), L8 and L10 (p = 0.06), and L9 and L10 (p = 0.53) were all greater than 0.05. The p-values of ADNs between levels other than these were less than 0.05. Combined with the results of the one-way ANOVA between groups, in general, there were significant differences in the driver’s attention to the surrounding road conditions and environment. There was no significant difference in the drivers’ attention to road conditions when the volume of information increased by 20 bits after reaching L3. However, a significant difference in the drivers’ attention to road conditions was observed when the cumulative increase in the volume of information reached 120 bits (ADNs comparison between L3 and L9). The ADNs for different levels of information volume are shown in Figure 11.

The graph shows that, in general, the ADN decreases as the volume of information increases, with two fluctuations: first an increase and then a decrease. This suggests that, overall, as the volume of information increases, the time drivers spend attending to environmental road conditions, in addition to signs, gradually decreases, indicating a reduction in driver distraction. However, when the information volume is at L1 and L2, the driver’s ADN exceeds 4.70 s, which are the two highest levels of ADN among all sign groups. This indicates that too little information can ensure that drivers attend to road conditions. The first rapid decrease in ADN occurs when the information volume is greater than L2. When reaching L4, the first trough of attention time to road condition occurs, indicating that when the volume of information is low (at L1 to L4), drivers gradually decrease their attention to road conditions while maintaining a certain level of attention. When the information volume goes from L4 to L6, there is a rebound in ADN. Based on the unstructured interviews conducted after the experiment, drivers become more attentive to the road condition ahead as the information volume continues to increase, and they actively allocate their energy to pay attention to the road condition. When the information level exceeds L6, the driver’s attention time to the road condition continues to decrease, with a significant decrease in ADN observed as the information level goes from L8 to L9 (p = 0.01). At L10, ADN reaches its minimum value. This phenomenon can be explained well by the unstructured interviews conducted after the experiment. It was found that the signs conveyed more information than the driver could process, and drivers could not divert more attention to road conditions when the information volume exceeded L8. However, too little attention to road conditions could lead to increased driving risks.

The results of this study present the same qualitative results as an existing study by [90]. However, the duration of attention to road conditions was somewhat shorter compared to the traffic tunnel. However, no mandatory violation penalties were set up in underground construction caverns compared with the transportation tunnels. In this case, drivers still shortened their attention to the roadway. This indicates that the underground construction environment is more complex, and drivers have more urgent needs for signs. Drivers expect more guidance from signs, thus reducing their attention to the roadway.

4.2.3. Recognition Rate of the Sign Group

Using the extracted NCA and NA, the value of RRSG was calculated. RRSG is the ratio of NAC to NA, which represents drivers’ ability to capture information. The higher the RRSG, the stronger the information capture ability of drivers. Multiple comparisons of RRSG at different levels of information volume were performed. It was found that the p-values of RRSGs between L1 and L2 (p = 0.06), L2 and L3 (p = 0.17), L3 and L4 (p = 0.58), L3 and L4 (p = 0.58), L4 and L6 (p = 0.06), L3 and L4 (p = 0.58), L5 and L6 (p = 0.60), L6 and L7 (p = 0.33), L6 and L8 (p = 0.20), L6 and L9 (p = 0.08), L7 and L8 (p = 0.72), L7 and L9 (p = 0.38), L7 and L10 (p = 0.24), L8 and L9 (p = 0.65), L8 and L10 (p = 0.44), and L9 and L10 (p = 0.66) were all greater than 0.05. The p-values for the RRSGs between levels other than these were less than 0.05. Combined with the results of the one-way ANOVA between groups, in general, there were significant differences in drivers’ ability to capture information. However, this variability needs to be increased by every two levels, namely 40 bits of information, before it becomes apparent. Meanwhile, when the volume of information exceeded L6, there was no significant difference in the recognition rate. The RRSGs for different levels of information volume are shown in Figure 12.

As can be seen in Figure 12, the RRSG appears to show an overall decreasing trend as the information volume increases. This indicates that, in general, as the information volume increases, the recognition rate of signs gradually decreases, and the driver’s ability to capture information decreases. Among them, when the information volume is less than L4, the recognition rate can still reach more than 95%, and when the information volume goes from L4 to L5, the efficiency of the driver’s recognition of signs extremely decreases. When the information volume reaches L7, the efficiency of the driver’s recognition rate drops sharply, with recognition efficiency below 85%. This indicates that when the driver’s recognition rate of information on the sign is too low, the driver is unable to capture the information in the sign group completely, and the information provided on the sign cannot be transmitted completely and effectively. Thus, the driver cannot make a driving decision correctly, and the driving risk surges.

The results of this study are the same as the qualitative results of existing visual cognition studies in tunnel driving [54]. However, quantitative exploration of sign recognition rates revealed that drivers have more difficulty recognizing signs in underground construction caverns. This is because underground caverns are still in the construction stage, where traffic and safety signs are staggered and appear in clusters over short distances without being optimized. Therefore, there is a need to improve drivers’ ability to capture and recognize signs by rationally reducing the volume of information in sign groups.

4.2.4. Processing and Cognitive Ability of Information

ADA reflects the time drivers spend processing information. The higher the ADA, the lower the processing and cognitive ability of the information. Multiple comparisons between ADA at different levels of information volume were performed. It was found that the p-values of ADAs between L3 and L6 (p = 0.10), L4 and L5 (p = 0.08), L4 and L8 (p = 0.63), L5 and L8 (p = 0.48), L6 and L7 (p = 0.60), and L9 and L10 (p = 0.29) were all greater than 0.05. The p-values of ADA between levels other than these were less than 0.05. Combined with the results of the one-way ANOVA between groups, in general, there were significant differences in drivers’ information processing and cognitive abilities across information environments. However, at lower information levels (L3, L4, and L5) and higher information levels (L6 and L7), drivers may exhibit the same information processing ability. The ADAs for different levels of information volume are shown in Figure 13.

As can be seen from the above graph, as the information volume increases, the ADA appears to increase and then decrease twice, with an overall upward trend. This indicates that, in general, as the information volume increases, the time spent looking at the sign increases, the time spent processing information grows, the ability to process and perceive information decreases, and the efficiency of information processing decreases. At the lowest information volume (L1), ADA is minimized, indicating that the efficiency of information processing is maximized when the lowest information volume is available; here, the driver can not only process the sign information quickly but also will pay sufficient attention to the other road conditions and environmental information that arrives. When the information volume is low (L1–L5), the ADA gradually increases. This indicates that an increase in the information volume enhances the driver’s attention to the sign, and the processing ability of the information conveyed by the sign subsequently decreases. And when the information volume reaches L6, the driver’s continuous gaze time on the sign decreases. This indicates that the driver realizes the increase in the information conveyed by the sign, actively speeds up the search of the sign information and improves the cognitive efficiency of the information processing and processing to adapt to the information volume at this time. When the information volume exceeds L6, the driver’s continuous gaze time on the signs increases again. This indicates that as the information volume in the sign group continues to increase, the driver’s attention to the signs continues to increase, and the efficiency of information processing continues to decrease. Drivers cannot process the sign information quickly and also do not pay enough attention to the other road conditions and environmental information that arrive. At the same time, combined with Figure 12, it can be seen that when the information volume exceeds L6, as the information volume in the sign group increases, the driver spends more time gazing at the sign group, but the driver still cannot increase the recognition rate of the signs. This indicates that there is a risk of information overload when the information volume in the sign group exceeds L6.

Existing related studies explore the efficiency of information processing from the perspective of the average duration of gaze behavior [27]. However, the existing studies exploring information processing capability from the perspective of average gaze duration are not comprehensive. Recognition of markings is the prerequisite for information processing capability. A driver can only process the information conveyed by a sign if the sign is first captured visually. This study begins with visual cognitive processes, combining attention allocated to roadway attention, sign recognition, and information processing of signs. It helps to explain and complement each other between different evaluation indicators under the influence of the amount of information.

4.2.5. Mental Load

When the driver tries to look at a target object, the degree of effort and psychological load is larger, the visual load is enhanced, and the pupil diameter will increase. A smaller pupil diameter can represent a driver’s smaller mental load, less psychological pressure, and more concentrated attention due to changes in the external environment and can also represent extreme mental fatigue caused by changes in the information volume of the external visual stimulus. Aiming at identifying the change in the driver’s pupil diameter caused by the change in the external environment, the normal test value of a driver can be compared and analyzed. Hence, multiple comparisons between MDPs at different levels of information volume were performed using a one-way ANOVA.

It was found that the p-values of MDPs between L1 and L3 (p = 0.09), L1 and L4 (p = 0.19), L1 and L5 (p = 0.11), L1 and L7 (p = 0.31), L1 and L8 (p = 0.26), L1 and L9 (p = 0.29), L2 and L3 (p = 0.17), L2 and L7 (p = 0.08), L2 and L8 (p = 0.06), L2 and L9 (p = 0.95), L3 and L9 (p = 0.28), L3 and L10 (p = 0.56), L4 and L5 (p = 0.63), L4 and L6 (p = 0.13), L4 and L7 (p = 0.79), L4 and L8 (p = 0.71), L5 and L6 (p = 0.23), L5 and L7 (p = 0.94), L5 and L8 (p = 0.98), L6 and L7 (p = 0.26), L6 and L8 (p = 0.29), L7 and L8 (p = 0.93), L7 and L9 (p = 0.11), L8 and L9 (p = 0.09), and L9 and L10 (p = 0.11) were all greater than 0.05. The p-values for MDPs between levels other than these were less than 0.05. Combined with the results of the one-way ANOVA between groups, in general, there were significant differences in the driver’s mental load. A significant difference was only observed when the increment of information volume exceeded 40 bits. Meanwhile, when the volume of information exceeded 100 bits (L5), the difference in the psychological burden on the driver per 40 bits increment was no longer significant. In addition, the p-values of MDPs between L3 and L6 (p = 0.00) and L6 and L10 (p = 0.00) were less than 0.05. The variation characteristics in the MDP of drivers in different information environments are shown in Figure 14.

As can be seen in the above figure, during driving in the underground cavern, the driver’s mean pupil diameter showed an increasing and then decreasing trend with the volume of information and then increased again. This indicates that the MDP gradually increases when the volume of information is at a lower stage (L1–L3). This indicates that drivers are more nervous about approaching signs, their need to perceive and judge the relevant traffic information requires more effort, the mental load is higher, and the mean pupil diameter is larger. As the volume of information continues to increase (L3–L6), the MDP decreases, and a trough occurs. The driver gradually relaxes physically and mentally, the mental load decreases, and the mean pupil diameter gradually decreases. The minimum value was reached in the L6 information environment, which indicates that in the L6 information environment, the driver was the most relaxed, with the lowest mental load and the smallest pupil diameter. As the volume of information continued to grow, the MDP gradually increased and reached a peak in the L10 information environment. This indicates that the average pupil diameter gradually increases when the information volume exceeds L6, and an excessive information volume causes the driver’s psychological load to increase.

This result differs somewhat from existing studies. Most of the existing studies on visual loads of tunnel driving focus on studies related to tunnel entrances and exits and smooth driving sections [35]. However, the results of simulated driving tests in underground construction tunnel environments are similar to those of visual load studies on highway driving because drivers do not need to rely on signs to guide them to intersections in traffic tunnels. In underground construction chambers, drivers need to rely on signs to ensure that they are driving on the correct side roads and avoiding construction risks. This is a little similar to the role of signs on highways.

4.3. Comprehensive Evaluation of Visual Cognitive Effects

4.3.1. Principal Component Analysis of Evaluation Indicators

This paper conducted KMO and Bartlett’s sphericity tests to assess the suitability of visual recognition evaluation indicators for factor analysis, as listed in

Table 7. KMO, Bartlett’s, and communalities test results.

Communalities Test			KMO and Bartlett’s Test
Evaluation Indicators after Standardized Conversion	Initial	Extraction	KMO	df	Sig.
ZCRC	1.000	0.775	0.618	6	0.000
ZRRSG	1.000	0.752
ZPCAI	1.000	0.842
ZMDP	1.000	0.968

Extraction method: principal component analysis.

. The KMO results (KMO = 0.62 > 0.60) indicated moderate suitability for factor analysis, while Bartlett’s sphericity test (Sig. = 0.00 < 0.05) suggested the presence of a correlation between the variables and that the data were suitable for factor analysis.

Meanwhile, an initial eigenvalue and a principal component score coefficient matrix were generated, as shown in

Table 8. Principal component calculation.

Component	Initial Eigenvalues			Component Score Coefficient Matrix
	Total	% of Variance	Cumulative %	ZCRC	ZRRSG	ZPCAI	ZML
1	2.250	56.247	56.247	0.37	0.385	0.400	0.012
2	1.087	27.179	83.425	−0.224	−0.003	0.227	0.904
3	0.399	9.968	93.393
4	0.264	6.607	100.000

. The result shows the eigenvalues of the first two principal components are greater than 1.000, which satisfies the selection criteria for principal components. The cumulative contribution rate of these two principal components is 83.425%, which is greater than the threshold of 60%, indicating that these two components can effectively represent the original data.

Table 8 indicates that the component score coefficients for each index in principal component 1 are 0.37, 0.385, 0.400, and 0.01. The component score coefficients for each indicator in principal component 2 are −0.224, −0.003, 0.227, and 0.904. Furthermore, the ratios of the variance contribution percentages of the two main components to the cumulative total contribution percentages are 0.674 and 0.326, respectively.

4.3.2. Drivers’ Visual Cognitive Intensity

We used Equations (8) and (9) to quantify the F-values under varying information quantities, and then a comprehensive evaluation of the visual cognitive effects while driving during underground tunnel construction was conducted, as demonstrated in Figure 15.

As can be seen from the figure, the driver’s overall visual cognitive efficiency decreases significantly as the volume of information increases. However, after the level of information volume reaches L6, the visual cognitive efficiency stabilizes. It suggests that when the volume of information exceeds L6, a driver cannot improve cognitive efficiency even with self-regulation and driving safety cannot be ensured. Therefore, efforts should be made to keep the information content in sign groups in underground construction caverns below 120 bits to ensure the visual cognitive effect of drivers.

Existing related studies inadequately consider the influence of visual cognitive processes on the effects of integrated visual recognition of signs. We found that when evaluative indicators were identified using visual cognitive processes and weights were given to the evaluative indicators, the results obtained appeared to correlate with the results of the post-test unstructured interviews. Simultaneously, controlling the volume of information conveyed by sign groups within a reasonable range is the key to optimizing sign layout and enhancing the comprehensive efficiency of visual cognition.

4.4. Implications

To the best of our knowledge, this paper represents a pioneering endeavor to systematically investigate and quantify the information conveyed by intricate sign groups within underground construction environments. Additionally, it aims to comprehensively analyze the impact of drivers’ visual perception of sign groups under various information loads, utilizing the framework of visual cognition and information theory. The ultimate objective is to propose a research framework for optimizing the layout of sign groups. Previous studies have explored the application of visual cognition in the optimization of traffic sign layout, including visual capture of signs and information transmission of signs [9,75]. Although these studies explored the effectiveness and usability of a single-sign layout, there is still a gap in the literature on driver visual perception of sign groups composed of safety and traffic signs and the optimization of sign group layout in the complex cave environment of underground structures. Thus, this study examined the comprehensive effect of driver visual cognition of sign groups under different information levels in underground construction caverns and proposed suggestions for optimizing the layout of sign groups.

Firstly, this study introduces a quantitative method to enhance the visual communication of sign group information in underground construction areas. However, previous research primarily focused on the basic elements of traffic signs to develop models quantifying sign information [27,91]. The aim is to assess the impact of individual traffic signs, with varying information quantities, on driver visibility in operational traffic tunnels. Signs within underground construction caverns are primarily arranged in groups, encompassing safety signs as well. The initial sign information model, constructed solely using the fundamental elements of traffic signs, is no longer valid. This study integrated the fundamental features of both traffic and safety signs to develop a comprehensive model for quantifying sign group information. This research establishes a quantitative foundation for the visual communication of sign groups in the underground construction environment.

Secondly, this study uncovers the visual cognitive process of drivers, develops a visual cognitive model, and devises an evaluation method to comprehensively assess the impact of visual cognition. This research offers a foundation for evaluating the overall visual cognitive impact of drivers on the information conveyed by complex sign groups while driving in construction areas. Previous studies on evaluating visual cognition of signs have primarily concentrated on analyzing traditional eye movement parameters, while the development of systematic evaluation indicators and comprehensive assessment research has received limited attention [92,93]. This study quantifies the visual cognitive processes of construction vehicle drivers for project managers and scholars, which can provide a basis for the comprehensive assessment of driver visual recognition efficiency.

Finally, this study provides empirical evidence for the visual cognitive effects of sign groups with different amounts of information on drivers in underground construction caverns. The results of this study can help project managers and researchers better understand the visual preferences and cognitive effects of engineering vehicle drivers and further optimize the layout of sign groups from a human factor perspective.

5. Conclusions

This study aims to explore the optimization of sign layouts in underground construction sites. This paper developed a calculation method for sign group information volume and a visual cognition model to determine a comprehensive evaluation method for sign group visual cognition. The comprehensive visual cognition of sign groups with different information volumes was evaluated using eye-movement data required for collecting indexes using eye-tracking, and suggestions for layout optimization were made using the results.

Low information volume does not improve recognition ability or alleviate the psychological burden. Excessive information can lead to missing signs on the left and top. Furthermore, with over 120 bits of information volume within 100 m, drivers cannot improve cognitive efficiency and driving safety even with self-regulation. A well-designed layout can help drivers process information quickly while having enough time to pay attention to road conditions and environmental information within the allowable cognitive load.

Overall, this study preliminarily tested the comprehensive visual cognitive effect of sign group layout by revealing the impact of underground cavern sign group information on the visual cognitive efficacy of drivers. This research has practical value in optimizing complex sign layouts in construction areas, which fills the gap in the original research on sign layout optimization in underground construction areas. This study will help to optimize the sign cluster layout in underground construction areas and further enhance the construction and transportation safety of sustainable hydropower plant construction.

Nevertheless, the current research has some limitations. This paper focuses only on the efficacy of visual cognition, and its reliability needs to be confirmed using the analysis of drivers’ decision-making behavior. In a follow-up study, the experiment will further explore the optimization of sign group layout combined with driver decision-making behavior.