Individual Behavior and Attention Distribution during Wayfinding for Emergency Shelter: An Eye-Tracking Study

Abstract

A fast evacuation from buildings to emergency shelters is necessary and important after the occurrence of a disaster. We investigated the variations in physical behaviors and cognition processes while finding emergency shelter. The on-site emergency-shelter-finding experiments were conducted in Beijing, China. Participants performed the task by using a wearable eye-tracking device. We aimed to assess three eye metrics: fixation counts, mean fixation duration, and visual attention index, to perform cognitive searching analysis for the environmental elements. The results showed that most people spend more fixation time on digital maps (297.77 ± 195.90 ms) and road conditions (239.43 ± 114.91 ms) than signs (150.90 ± 81.70 ms), buildings (153.44 ± 41.15 ms), and plants (170.11 ± 47.60 ms). Furthermore, most participants exhibit hesitation and retracing behaviors throughout the wayfinding process. The participants with relatively rich disaster experience and a proactive personality exhibit better performance in the shelter-finding task, such as a shorter retracing distance (p = 0.007) and nearer destination (p = 0.037). Eye metrics, together with the questionnaire, can mirror the complexity and heterogeneity of evacuation behavior during emergency shelter-finding. In addition, this also provides insights for the optimization of guidance sign systems and improvements in emergency management.

1. Introduction

1.1. Background

Urbanization and population growth have increased the exposure of communities to natural and man-made disasters in various regions worldwide [1]. According to the governmental statistics published by the Ministry of Emergency Management of the People’s Republic of China, 138 million people were affected by natural disasters in 2020, leading to economic losses of CNY 370.15 million [2]. Considering fire accidents, the rescue department received 252,000 alarms, among which community fires accounted for 43.3%. This led to long-term challenges for both governments and the public in densely built neighborhoods. Therefore, emergency shelters and open public spaces suggest a growing concern regarding the feasibility of providing places for rescue activities and temporary accommodation.

There still exist some issues regarding the construction and service level of emergency shelters to enhance the integration of disaster response with sustainable urban planning. In light of our previous research and onsite surveys, the main problems in China’s emergency shelters include a lack of publicity and preparedness, incomplete ancillary facilities, and an inadequate service level [3]. For instance, the current guidance system in China is poorly effective when used as an aid for evacuation. Galea et al. have shown that only 38% of people are able to see and understand emergency signs in presumed dangerous situations [4]. Therefore, one important element for developing high-quality emergency shelters is the presence of a scientific guidance system to enable residents and tourists to effectively and safely move towards open spaces which are specifically designed to accommodate the people living or working in the surrounding communities.

1.2. Literature Review

In order to further analyze wayfinding behavior, both in buildings and open public areas, most of the current research tends to focus on conducting surveys, video analyses, or simulations to provide a good approximation of the real world [5]. For example, traditional video analysis demonstrates that, in large complex environments such as hospitals, signs alone are not sufficient for wayfinding, and the presence of an extra entrance is more effective for visitors and patients [6,7]. Questionnaires are another traditional research method [8,9]. For example, at train stations, one questionnaire demonstrated that, although the majority of passengers were familiar with the environment, only 43.8% of people could see the emergency exit signs [10]. A questionnaire was also designed to investigate passengers’ perception of wayfinding in two different airports [11]. According to the survey conducted by N. Shiwakoti et al., there was a negative association in terms of feeling safe at the airport with the performance of emergency wayfinding. Simulations developed based on Pathfinder and Anylogic were also adopted to examine the evacuation situation in a subway station [12,13], an underground shopping mall [14], and a high-rise building [15]. Different measures have been proposed to relieve evacuation pressure, though, for example, the effect of exit width has been found to be different in different research works. In addition, the wayfinding behavior of bike users has also been analyzed. The size of the pedestrian crowd, the opinions of bike users, and the experience with the wayfinding signage were investigated qualitatively [16]. Later, visibility models based on signage and direction helped 3D systems to identify salient landmarks and assist in wayfinding [17]. With the development of virtual environment technology, realistic scenarios such as garden mazes, office buildings, and subway stations have been developed for testing the wayfinding inclinations and visuospatial abilities of the subjects [18,19,20]. The emergence of physiology signal measurement provides new approaches for studying human behaviors. For large transportation hubs, scholars have conducted physiological and cognitive load monitoring and analysis on transfer passengers, finding that the individual’s level of environmental perception can be obtained through measures such as skin conductance, heart rate, and eye movement [21]. Scholars have also used eye-tracking devices to record individuals’ behavior in terms of gaze towards the surrounding crowd when searching for safe exits in buildings, revealing the influence of external factors on route selection and attention allocation [22]. Others have also conducted feature clustering analysis on visual scanning paths related to individual search behavior [23], providing new methods for evaluating evacuation behavior and performing psychological analysis.

The interaction between escape signs and pedestrian behavior has attracted a lot of attention from researchers. Few researchers have studied the decision process and behavior after pedestrians have seen the signs [24,25]. A cellular automata (CA) model was applied to analyze the impact of evacuation signs on evacuees’ movements by introducing a diagonal-moving parameter and waiting-urgency factor [26]. In addition, a piece-wise probability function was employed by Zhang et al. to describe the interactions between pedestrians and evacuation signs [27]. Furthermore, computer graphics were used to evaluate route signs [28]. After the 2011 tsunami, the location of tsunami-related signs has been taken into consideration by the government [29]. P. Ma investigated the escape time with different view radii of signs in a building based on the Vicsek model [30]. Kubota et al. conducted an analysis of the installation position and arrow type of emergency evacuation signs in a complex layout based on the validity and confidence of the subjects during wayfinding [31]. Additionally, forbidden signals, which indicate the prohibition of using a certain exit, have been analyzed and designed [32]. Furthermore, the characteristics of variable message signs for road tunnel emergency evacuation were evaluated using a questionnaire and eye-tracking device [33]. It is noteworthy that the green pictogram on the signs was found to be ineffective for providing navigation for drivers [32,34]. On the contrary, a green background is recommended for dissuasive signals [32]. Scholars have also used physiology signal measurements to optimize the design of evacuation signage in various settings, such as high-rise hotels [35] and office [33]. Their focus primarily revolves around investigating the relationship between personnel evacuation behavior and sign design [36]. This encompasses aspects including sign location, color, graphics, flashing frequency, etc. Furthermore, they have examined attention distribution at different types of intersection, changes in heart rate and skin conductivity [37], as well as significant changes in electroencephalographic signals [38]. Their findings can serve as a foundation for the design of exit signage in complex structures. The research methods, scenarios and subjects related to wayfinding behavior and evacuation were investigated in prior studies, which can be categorized as shown in Figure 1.

After reviewing peer research, some key points are identified:

• Most of the works presented in the literature focused on wayfinding in various contexts such as building environments, underground spaces, highways, transportation systems, etc. However, the analysis of wayfinding behaviors for emergency shelters in urban environments remains a major challenge and has received limited attention. In the event of a natural or man-made disaster, the efficient evacuation of individuals from building areas to outdoor emergency shelters is crucial for ensuring their safety and preserving lives.
• The aforementioned research works have primarily utilized questionnaires, simulative models, video analysis, and VR techniques. These approaches have been employed to optimize the effectiveness of indoor evacuation signs through sign design improvement, determining optimal installation positions, and optimizing the displayed contents. However, it is noteworthy that the attention distribution and cognitive search process during emergency-shelter finding are rarely analyzed. Therefore, wearable equipment, such as eye-tracking glasses that capture the attention distribution of individuals, can be incorporated and combined with on-site experiments to evaluate the efficacy of directing signs for emergency shelters.

1.3. Contribution

To solve the above shortcomings in research, on-site experiments were conducted to examine the physical behavior and cognition processes during wayfinding towards emergency shelters. The main contributions of this study are as follows: (1) the utilization of an eye-tracking device to capture the attention distribution and cognition processes of the evacuee. This approach supersedes traditional methods such as video analysis, VR techniques and simulations. The eye metrics obtained from the eye-tracking devices provide a clear understanding of the interaction between the surrounding elements and evacuees. (2) The on-site experiment allowed for the differentiation and identification of the complexity and heterogeneity of individual evacuation behavior. This study demonstrated the influence of demographics and background factors on shelter-finding behaviors. These findings contribute to the enhancement of current simulative models and promote the more realistic simulation of evacuation scenarios. (3) This study represents a pioneering attempt to investigate the behavior and attention distribution of evacuees during the emergency-shelter-finding process. The results offer useful clues for the design of emergency shelter signs and the organization of public drills. Ultimately, these findings will contribute to the improvement of emergency management and response, with a particular focus on urban safety considerations.

2. Materials and Methods

2.1. Experimental Design

In the first phase, we conducted a survey of Beijing and identified a list of potential districts for analysis. Then, our focus narrowed down to the Yuan Dynasty City Wall Relics Park, which offers a vast open space with a long strip of land. This park serves as the largest emergency shelter established based on the concept of open space. According to the land use data obtained for building usage, there were residential, educational, commercial, and office buildings in close proximity to the park. Yuan et al. estimated the daytime and nighttime population size by using a 500 m × 500 m grid and observed consistently high population densities near the park, ranging from 5000 to 20,000 people. Therefore, the demand for emergency shelters in the event of any disaster was significantly high [34].

A total of 17 voluntary participants took part in this study, which was conducted in the spring of 2021. All individuals were college students aged between 19 and 23, and they were in good physical and mental health. The demographic information of the participants is presented in

Table 1. The demographic information of the participants.

Demographic Characteristics	Classification	Frequency	Percentage
Age	20 and below	3	17.6%
	21–30	14	82.4%
Gender	Male	10	58.8%
	Female	7	41.2%
Education level	Undergraduate	7	41.2%
	Graduate	10	58.8%

. None of the participants were familiar with the study region. Prior to the wayfinding task, participants were asked about their basic information, safety awareness and any experience with disasters. Then, all participants were instructed to wear mobile eye tracking glasses and complete the wayfinding task. As depicted in Figure 2, all participants were instructed to begin their assignment individually from the same location. They were informed that they were currently located at the entrance of the Mudanyuan community and needed to proceed to the nearest emergency shelter promptly. There were several emergency shelters located around the community, and participants had the freedom to choose any destination they preferred. It is important to note that the participants were permitted to use a smartphone navigation app during the wayfinding task. After finishing the assignment, participants were asked to recall their wayfinding behavior throughout the experiment. The experimental protocol was approved by the Academic Ethics Committee of the University of Science & Technology Beijing.

The eye metrics and subconscious behavior were recorded by using the Tobii Pro Glasses 2 (Tobii Technology, Danderyd, Sweden), a wearable eye-tracking device. These glasses recorded eye-movements at a rate of 120 Hz, with data transmitted wirelessly via a portable unit. The glasses were equipped with four internal and external digital cameras, providing a resolution of 1920 × 1080 at 25 fps. This allowed us to acquire high-definition images for gaze heatmaps. The color scheme of the heatmap depended on the duration of the subject’s gaze on a particular stimulus. Longer fixation times were represented by a transition from green to yellow, orange, and finally red. The glasses utilized corneal reflection technology and dark pupil tracking technology to ensure accurate eye movement data capture. Tobii studio software (version 1.79) was used to calibrate and record the eye metrics. A single circular calibration board was employed to correct the gaze point and pupil position for each participant. The entire process of recording eye-movement data for each participant lasted about 30 to 50 min. This encompassed equipment calibration, the wayfinding task for the emergency shelter, and questionnaire completion. Then, Tobii Pro Lab Analyzer Edition was used to analyze the gaze heatmaps and eye metrics.

The study incorporated several key factors, including demographics, background information, physical behaviors, and eye metrics, as outlined in

Table 2. The main variables collected and analyzed in the research.

Factors	Variables	Source
Demographics	Age, gender and education level	Pre-experiment Questionnaire
Background information	Safety awareness
	Evaluate the signs and recall the behavior	After-experiment Questionnaire
	Self-assessment
Physical behavior	Experiment duration, Proportion of hesitation, Proportion of detour, Extra route length ratio, Rotation of the electric map	Eye-tracking device
Eye metrics	Dwelling time, Fixation counts, Mean fixation duration and Visual attention index	Eye-tracking device

. Statistics analysis was conducted to investigate the impact of demographics and background factors on evacuees’ behavior and attention distribution. Independent-sample t-tests were employed to assess differences in a variable between two groups for continuous parametric data, while one-way ANOVA was used to determine the differences in a variable among three or more groups. For continuous nonparametric data, Mann–Whitney U-tests were employed to assess differences between two groups, while Kruskal–Wallis H-tests were used to determine differences in a variable among three or more groups. MATLAB 2019a was used for performing the statistical analysis, and findings were reported at a significance level of p < 0.05. To investigate the relationship between these variables, a correlation matrix was generated using data collected from the entire sample.

2.2. Eye Metrics

The objective of this study was to identify the observation patterns of participants as they searched for emergency shelters around the community. Before extracting eye metrics, dynamic areas of interest (AOIs) were identified based on frame-by-frame analysis. Five common objects that participants frequently observed during the experiment were designated as AOIs, including roads, buildings, directing signs, phones (electric map), and plants. The gaze numbers, gaze positions, and fixation duration of participants within the predefined AOIs were measured and recorded at regular time intervals. In this study, the average fixation duration during the wayfinding process was approximately 180 ms–275 ms. To determine if a fixation occurred, at least 22 consecutive frames were required to be counted as one fixation. The eye metrics used in this study are defined as follows:

Dwelling time (DT): The total time (ms) that participants spent fixating within a specific AOI. A longer dwelling time shows that participants are allocating a significant amount of visual attention to a certain object.

$${\mathrm{D}\mathrm{T}}_{\mathrm{j}}={\sum }_{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{E}\mathrm{T}}_{\mathrm{i}\mathrm{j}}-{\mathrm{S}\mathrm{T}}_{\mathrm{i}\mathrm{j}})$$

(1)

where ${\mathrm{D}\mathrm{T}}_{\mathrm{j}}$ denotes the dwell time of the jth AOI, and ${\mathrm{E}\mathrm{T}}_{\mathrm{i}\mathrm{j}}$ and ${\mathrm{S}\mathrm{T}}_{\mathrm{i}\mathrm{j}}$ represent the ending and starting times of the ith fixation of the jth AOI, respectively.

Fixation counts (FC): The total number of recorded gaze points within a defined AOI.

Mean fixation duration (MFD): The average value of the fixation duration (ms), calculated by dividing the total dwell time by the number of fixation counts.

$$\mathrm{M}\mathrm{F}\mathrm{D}=\frac{\mathrm{D}\mathrm{T}}{\mathrm{F}\mathrm{C}}$$

(2)

Visual attention index (VAI): This index is defined as the proportion of the total fixation duration relative to the time spent in saccades. A smaller VAI value indicates that participants spent more time engaged in visual searching rather than recognition.

$${\mathrm{V}\mathrm{A}\mathrm{I}}_{\mathrm{j}}=\frac{{\mathrm{D}\mathrm{T}}_{\mathrm{j}}}{{\mathrm{S}\mathrm{T}}_{\mathrm{j}}}$$

(3)

where ${\mathrm{V}\mathrm{A}\mathrm{I}}_{\mathrm{j}}$ denotes the visual attention index for jth AOI and ${\mathrm{S}\mathrm{T}}_{\mathrm{j}}$ is time (ms) spent in saccades.

3. Results

3.1. Participants’ Safety Awareness

The questionnaire consisted of three parts: safety awareness and experience, evaluation of the guidance signs, and recalling the shelter-finding experiment (

Table 3. A summary of safety awareness and wayfinding behaviors of subjects.

No.	Variables	Disagree	Neutral	Agree
Part A	Safety awareness and experience
1	Have experience of real disaster	61%	0%	39%
2	Familiar with emergency shelter	11%	89%	0%
3	Knew the location of nearest emergency shelter near the home or company	22%	50%	28%
4	Usually noticed various guiding signs on the street	17%	50%	33%
Part B	Evaluate the guiding signs and recall the shelter-finding experiment
5	Saw the guiding signs for emergency shelter in the experiment	6%	0%	94%
6	The signs provide useful information for wayfinding task	6%	44%	50%
7	Understand the content of the guiding signs	17%	17%	66%
8	Prefer to finish the wayfinding task by using electric map rather than ask people	11%	0%	89%
9	The electric map in the phone was more helpful than guiding sign	0%	72%	28%

). In addition, it included a self-assessment section based on a hypothetical scenario, as shown in

Table 4. A summary of self-assessment of subjects in hypothetical disaster scenario.

No.	Variables	Response
Part C	Self-assessment in hypothetical scenario
10	Possible psychological response when facing disaster	“Calm down and observe surroundings”	44%
		“Feel scared and would escape firstly”	17%
		“Feel confused and would evacuate with other people”	39%
11	Possible physical behavior when facing disaster	“Would not take any actions at first”	23%
		“Try to explore the possible route by myself”	33%
		“Keep up with surrounding people”	44%
12	How to plan the evacuation route	“Follow the crowd”	33%
		“Follow the guiding signs”	67%

. Part A comprised four questions aimed at understanding the participants’ emergency experience and familiarizing them with the emergency shelters and guidance signs. Part B was administered to participants after the experiment to evaluate the existing guidance signs and recall their wayfinding behaviors for emergency shelters. Additionally, Part C focused on surveying possible psychological responses and physical behaviors when facing disasters.

Combining the eye-movement data with the questionnaire results provides insights into the subjects’ perception of the surrounding objects and their wayfinding strategies. In this study, although most subjects (39%) reported having experience with natural or man-made disasters, only 28% of the participants were aware of the specific locations of nearby emergency shelters. Furthermore, only 33% of the participants state that they paid attention to various guidance signs on the street. Interestingly, the majority of individuals (89%) had heard about the establishment of emergency shelters but lacked knowledge of their distribution and function. During the experiment, while most people (94%) noticed the guidance signs for emergency shelters, only 50% of the participants believed that the signs could provide useful information during the wayfinding task. In fact, 17% of the participants failed to understand the content of the guidance signs due to poor compliance between the sign direction and complex intersections. Despite the lack of continuity in the guidance signs, 72% of the participants considered them to be as important as digital maps when it came to the shelter-finding task. Furthermore, 89% of the participants preferred using digital maps to complete the wayfinding task rather than seeking assistance from others. Regarding their possible response to a sudden disaster, 17% of the participants reported feeling panicked. More than a third of the participants mentioned that they tend to observe and follow the actions of surrounding individuals, while 67% preferred to rely on the guidance signs during the evacuation process.

3.2. Physical Behavior in Wayfinding Task

The route trajectories and physical behaviors were observed and quantified through data collected from wearable devices. This allowed for the systematic quantification of hesitation and retracing behavior during evacuation. Although all participants started from the same location to find the nearby emergency shelter, they were given the freedom to choose different routes leading to various emergency shelters. Figure 3 shows the cluster of evacuation trajectories for all the subjects. It is important to note that while this task could be completed by selecting a convenient and short route, significant variances in the trajectories are observed in this study. We observed that approximately 11 out of 17 subjects reached a similar emergency shelter located west of the starting point, while the remaining participants reached emergency shelters situated further north, east, and south. Among these 17 subjects, participants a–c successfully arrived at their desired destinations without any route modifications. Figure 3d–k illustrate the route trajectories of eight participants who reached their destinations with one instance of retracing. Participants l–q spent a considerable amount of time and physical energy on detours and experienced difficulties in finding the emergency shelter smoothly compared to other people.

In addition, a detailed analysis of behaviors, including retracing and hesitation, during the shelter-finding process was conducted. Despite being able to follow signs and use navigation applications, all subjects exhibited instances of hesitation and retracing. In particular, the normal forward movement without hesitation or retracing indicated that participants feel relatively relaxed and confident in their decisions. Through the analysis of replayed eye-tracking videos, it was observed that participants would stop walking and look around, which was marked as a hesitation behavior. Hesitation behavior indicated that subjects spent more dwelling time on the surrounding objects. Furthermore, participants would retrace their steps when they realized they were heading in the wrong direction. It is evident from Figure 4a and

Table 5. The statistics of physical behavior of the participants.

Participants	Experiment Duration (min)	Proportion of Hesitation (%)	Proportion of Detour (%)	Extra Route Length Ratio	Rotation of the Electric Map
a	11.05	3.92	0	0.81	1
b	14.05	0.73	0	0.00	3
c	16.73	2.21	0	0.71	1
d	27.83	19.18	19.8	2.47	8
e	10.05	2.76	9.32	0.56	0
f	13.88	6.99	8.33	0.55	2
g	25.18	12.7	5.11	1.56	6
h	16.90	10.91	0.45	1.25	1
i	37.47	11.33	6.11	3.87	7
j	17.77	4.73	3.12	1.07	8
k	12.22	21.38	2.06	1.23	0
l	11.37	12.08	17.98	0.39	3
m	10.22	9.48	9.54	1.31	2
n	48.72	24.58	26.27	3.32	8
o	24.73	3.58	4.89	2.25	1
p	55.08	8.03	25.84	4.62	5
q	37.75	23.03	31.57	3.78	4
Mean	23.00	10.45	10.02	1.75	3.53
SD	14.01	7.60	10.34	1.39	2.92

that the time proportions of normal forward behavior, hesitation behavior, and retracing behavior were about 79.53%, 10.45%, and 10.02%, respectively.

To evaluate the wayfinding performance, an additional metric called the extra route length ratio is adopted. This ratio is calculated by dividing the extra route length (defined as the actual traveled length minus the measured shortest length) by the shortest length required to reach the emergency shelter. The extra route length ratio ranged from 0 to 4.62, as shown in Table 5. This clearly shows that most route modification behaviors occurred during the first half of the experiment for participants d–m (10 participants). The average time proportions of hesitation behavior and retracing were about 11.15% and 8.18%, respectively. Additionally, participants a–c exhibited few instances of hesitation behaviors when facing crossroads. It can be assumed that these participants were able to recognize the geometric characteristics of the environment from the beginning. However, participants n–q displayed intensive retracing and hesitation behaviors throughout the entire wayfinding process. The duration of the experiment for these participants was about 2.4 times longer than the others and included fewer retracing behaviors. The average time proportions of hesitation behavior and retracing were approximately 14.81% and 22.14%, respectively. Analysis of Figure 4a reveals that most route modifications occurred after hesitation behavior was observed.

The typical locations where subjects exhibited the most retracing behavior during the wayfinding process are illustrated in Figure 4b–e. For instance, some participants initially headed towards residential areas but retraced their steps upon realizing that emergency shelters are not located within densely populated communities. Furthermore, modifications in heading directions frequently occurred around pedestrian overpasses and pedestrian crossings. Participants also exhibited spontaneous doubt regarding their direction when encountering signs.

We conducted a series of statistical analyses to explore the influence of background factors on the physical behaviors of the participants. The computed results with significant differences are summarized and reported in

Table 6. The influence of background factors on the physical behavior of the participants was examined using t-tests, one-way ANOVA, and the Kruskal–Wallis H-test.

Physical Behavior	Variable	M ± SD	p	Mean Difference	95% CI
	Gender
Total hesitation duration (s)	Male	23.17 ± 13.16	0.045	−13.89	−27.43, −0.35
	Female	37.06 ± 12.48
Total detour duration (s)	Male	23.04 ± 23.90	0.034	−105.50	−200.02, −10.97
	Female	128.54 ± 102.04
Total experiment duration (min)	Male	16.48 ± 8.62	0.038	−15.83	−30.55, −1.12
	Female	32.32 ± 15.49
Rotation of the electric map	Male	1.90 ± 2.08	0.002	−3.9	−6.26, −1.66
	Female	5.86 ± 2.34
	Education *
Extra route length ratio	G	0.79 ± 0.43	0.001	−2.34	−3.33, −1.34
	UG	3.12 ± 1.08
Total hesitation duration (s)	G	22.64 ± 12.75	0.026	−15.19	−28.30, −2.09
	UG	37.83 ± 12.07
Trajectory modification.	G	0.90 ± 0.74	0.037	−2.39	−4.58, −0.19
	UG	3.29 ± 2.36
Total experiment duration (min)	G	13.42 ± 2.90	0.002	−23.26	−34.21, −12.30
	UG	36.68 ± 11.82
Rotation of the electric map	G	2.10 ± 2.33	0.010	−3.47	−6.00, −0.94
	UG	5.57 ± 2.50
Physical behavior	Variable	M ± SD	p	F	Mean square **
	Action ***
Extra route length ratio	1	1.72 ± 1.22	0.007	7.11	7.82 1.10
	2	0.97 ± 0.78
	3	4.20 ± 0.59
Trajectory modification.	1	1.33 ± 1.23	0.001	11.53	19.22 1.67
	2	1.33 ± 1.36
	3	6.001 ± 0.41
Total experiment duration (min)	1	22.931 ± 3.26	0.013	6.02	725.93 120.67
	2	15.31 ± 5.14
	3	46.42 ± 12.25
Rotation of the electric map	1	4.89 ± 3.10	0.035	6.32	26.01 6.02
	2	1.17 ± 1.17
	3	4.50 ± 0.71
Physical behavior	Variable	Mean rank	p
	Background safety value
Destination	low	14.5	0.037
	high	8.32
	medium	6.00

* Education: G means graduates, and UG means undergraduates. ** Mean square: The first value is the between-group mean square and the second value is the within-group mean square. *** The action referred to in question 11 in the questionnaire: 1 represents “wouldn’t take any actions at first”, 2 represents “tried to explore the possible evacuation route by myself”, and 3 represents “contact with the surrounding people”.

. Independent sample t-statistics revealed significant differences in total hesitation duration (p = 0.045, mean difference = −13.89), detour duration (p = 0.034, mean difference = −105.50) and total duration of the experiment (p = 0.038, mean difference = −15.83) based on gender. This suggests that, due to differences in physical strength, female participants spent more time on walking and finding their destination. Furthermore, as the continuity of the guidance signs was not effective for wayfinding, participants tended to rely on navigation applications during the experiment. Analysis shows that participants subconsciously rotated their phones when they felt confused about spatial orientation. The average number of phone rotations was 5.86 for female participants and 1.90 for male participants during the entire experiment. Therefore, there was a significant difference in the amount of phone rotation between male and female participants (p = 0.002). These results align with previous studies where men tended to rely on spatial orientation and navigation skills to find suitable paths, while women preferred to gather and process knowledge for reaching their destination [39]. Please also note that male and female participants employed different cognitive strategies during the experiment.

Furthermore, it is apparent that education also plays a crucial role in shelter-finding behavior. Significant differences were observed in the extra route length ratio (p = 0.001, mean difference = −2.34), trajectory modifications (p = 0.037, mean difference = −2.39), total hesitation duration (p = 0.026, mean difference = −15.19), total experiment duration (p = 0.002, mean difference = −23.26), and rotation of digital maps (p = 0.010, mean difference = −3.47) between graduates and undergraduates. Overall, undergraduates failed to reach the emergency shelter as smoothly as graduates. Another demographic factor related to education is age, as all the graduates in this study were older than the undergraduates.

The impact of self-assessment on wayfinding behavior was also investigated. When participants were asked about possible physical behaviors when facing a disaster, three types of responses were identified: passive, proactive, and dependent. These response types reflected different characteristics in terms of participants’ behaviors. Proactive participants, for example, exhibited a smaller extra route length ratio compared to the other types of participants (p = 0.007). Moreover, proactive participants were less likely to rotate the digital map during the shelter-finding process (p = 0.007). On the contrary, participants who relied more on others showed a longer experiment duration (p = 0.013), a larger extra route length ratio (p = 0.007), and a higher number of retractions (p = 0.001). In addition, in order to assess the safety awareness and experience of participants’ wayfinding behaviors, the responses to questions 1–4 in the questionnaire were summarized as the background safety value. It is interesting that a p-value of 0.037 for the Kruskal–Wallis test is reported for destination selection among participants with different background safety values. Specifically, participants with relatively rich experience regarding disasters and emergency shelters tended to find the nearest emergency shelters, while participants with lower background safety values tended to reach shelters located further away.

3.3. Analysis of the Contents Extracted from Video

3.3.1. Participants’ Eye Metrics when Viewing Different Elements

In order to identify the attention distribution and visual focus variance of different participants during the emergency-shelter finding, eye metrics including the visual attention index (VAI) and mean fixation duration (MFD) were examined, and are presented in Figure 5. The analysis included data from only seven participants due to missing data caused by equipment malfunction. The results show that most people allocated more fixation time to the digital map (phone) and the road, instead of signs, buildings, and plants. Based on the VAI, two distinct attention distribution patterns were observed. One group showed a relatively higher average VAI for the digital map (0.025), with a mean fixation duration of 244.17 ms. The other group exhibited a higher VAI for the roads (0.053), with a mean fixation duration of 251.58 ms. The trends of the VAI for the phone were consistent with the MFD presented in Figure 5c. Combining the eye metrics with physical behaviors, it was observed that participants who focused more on the digital map (higher VAI) tended to have longer retracing times. On the other hand, the fixation characteristics of participants for the road varied. Some individuals frequently gazed at road conditions with a shorter fixation duration, while others exhibited more saccadic behaviors with longer fixation duration. Regarding the other surrounding objects, the VAI and MFD followed the order plants > signs > buildings, where “>” means “greater than”. The lower VAI and MFD values suggest that participants spent more time searching (or making saccades) when catching the sight of signs and buildings. However, the average VAI and MFD for plants increased overall, as some participants (e.g., l and q) exhibited different patterns of attention distribution. Most participants spent an average of 158.81 ms for each fixation point on plants, while participants l and q spent an average of 203.98 ms for each fixation point. This phenomenon can be attributed to participants searching for directing signs hidden among plants during the process of emergency-shelter finding.

Figure 6 shows the high dispersion of fixation counts (FCs) observed for different objects. The mean FCs for signs, road, buildings, digital map, and plants are 13, 114, 9, 86, and 44, respectively. There are significant differences in the number of fixations for roads, buildings, and phones between males and females. Specifically, female participants exhibit a higher FC (mean FC = 262 and SD = 106) compared to male participants (mean FC = 65 and SD = 69) for road conditions during the wayfinding process, with a statistical difference in FC between female and male participants (p = 0.019). In addition, female participants also show a higher FC compared to male participants for surrounding buildings along the road (p = 0.033). Overall, the total FC for surrounding objects during the entire experiment is different between females (mean FC = 437 and SD = 137) and males (mean FC = 164 and SD = 136), p = 0.002. These results indicate that female participants pay more attention to surrounding objects, especially roads and buildings. Furthermore, participants who reported in the questionnaire that they prefer to follow the surrounding people during the escape process (mean FC = 153 and SD = 131) tend to check the digital map on their phone more frequently, while those who claim to choose observing signs rather than following the crowd (mean FC = 46 and SD = 39) have fewer fixations on their phone. Individual visual behavior is depicted by the total fixations of participants, ranging from 38 to 639 during the entire experimental process. For instance, some participants (e.g., participants a and e) pay little attention to digital maps, while others (e.g., k and m) allocate more than 50% of their FC to their phone. Furthermore, all participants spend an average of 30% of total FC on observing road conditions. On the other hand, some participants (e.g., e and q) gaze at plants more frequently in order to find hidden directing signs along the road. Please note that buildings rarely attract as many fixations as other objects, demonstrating their weak role in guiding the participants.

In order to investigate the correlation among the eye metrics, demographics, and physical behaviors, a correlation matrix was created for the entire tested sample, as presented in

Table 7. The Pearson correlation coefficient (${\mathrm{R}}^2$) among eye metrics, demographics, and physical behaviors.

Indicator	Gender	Psychological Response	Total Experiment Duration	Rotation of the Electric Map	Extra Route Length Ratio	Hesitation Time	Total Detour Duration
Phone VAI	0.621	0.425	0.43	0.822 *	0.405	0.574	0.916 **
Road VAI	0.495	0.541	0.341	0.830 *	0.176	0.337	0.815 *
Phone MFD	0.569	0.48	0.41	0.806 *	0.339	0.511	0.851 **
Road MFD	0.657	0.298	0.456	0.752 *	0.528	0.643	0.904 **
Building MFD	0.641	0.736 *	0.659	0.59	0.568	0.354	0.479
Road FC	0.792 *	0.65	0.689	0.917 **	0.538	0.507	0.848 **
Building FC	0.748 *	0.786 *	0.689	0.838 **	0.5	0.359	0.732 *
Phone FC	0.932 **	0.52	0.883 **	0.770 *	0.880 **	0.751 *	0.729 *
Plant FC	0.684	0.507	0.832 *	0.295	0.777 *	0.392	0.081

* p < 0.05 (two-tailed), statistically significant at a significance level of 95%. ** p < 0.01 (two-tailed), statistically significant at a significance level of 95%. Pearson correlation: ${\mathrm{R}}^2$ ≤ 0.30 (weak); 0.30 < ${\mathrm{R}}^2$ < 0.70 (moderate); ${\mathrm{R}}^2$ ≥ 0.70 (strong).

. The results show a positive correlation between gender and fixation counts for roads, buildings, and phones. The phone rotation behavior and detour duration are strongly correlated with most eye metrics (${\mathrm{R}}^2$ > 0.73 and p < 0.05). A significant correlation is also observed between the duration of the complete experiment and FC for phones/plants (${\mathrm{R}}^2$ > 0.80). In particular, phone FC shows a positive correlation with all physical behaviors (p < 0.05). The reported psychological responses of the participants are also strongly correlated with the eye metrics for buildings. Participants who reported being calm in the face of danger exhibited fewer fixations and paid less attention to buildings, as seen in participants a and m.

3.3.2. Attention Distribution under Different Behavior

The visual attention distribution changes among different wayfinding behaviors suggest variation in the participants’ cognitive process. As presented in Figure 7, when participants move forward without hesitation, their gaze points are distributed horizontally along the road. When participants intend to modify their heading direction, their gaze points converge towards the end point of the path. The lateral and bottom parts of the visual scene are not the main focus of attention. When participants demonstrate hesitation, their wayfinding behavior can be classified into two styles: route-based and map-based, based on their gaze pattern. Specifically, map-based wayfinding is considered to be allocentric or coordinated [40]. Individuals pay more attention to the spatial relations and distances between themselves and the urban elements. They show dispersed gaze distribution around the digital map and road conditions, indicating the mental mapping process while navigating the environment. On the other hand, route-based wayfinding, reported in previous studies as egocentric orientation, involves participants directing more of their gaze towards surrounding objects, i.e., directing signs, in an effort to understand their surroundings. The heatmaps clearly demonstrate the different stimuli that capture participants’ attention based on their confidence level in their chosen paths.

3.4. Drawbacks of the Existing Evacuation Signs

The experimental results show that digital maps are more helpful than evacuation signs in wayfinding tasks, as supported by the qualitative analysis of eye metrics and physical behaviors collected during the experiment. The average FC, VAI, and MFD for directing signs are 13, 0.0045, and 150.90 ms, respectively, while the same eye metrics for digital maps are 86, 0.0338, and 195.90 ms, respectively. Considering the questionnaire, only 50% of the participants stated that the evacuation signs are useful for orientation., and 66% found guidance signs to be easy to understand. Therefore, most participants preferred to rely on digital maps to complete the wayfinding task. The outcomes extracted from the eye metrics revealed that participants spend fewer fixations on the directing signs during the process of shelter finding. However, the attention paid to some emergency shelter signs increased when participants experienced uncertainty in choosing the correct direction, particularly when the sign direction did not align well with the intersection layout.

The guidance signs for emergency shelters are shown in Figure 8. These signs are positioned at various locations around the community and provide information about the direction and distance to the nearby emergency shelters. However, the effectiveness of the existing signs is limited. For example, some signs are shielded by plants or temporary guardhouses, making them difficult for participants to spot. Furthermore, the arrows on certain signs pointing towards large intersections are not clearly understood by the evacuees. As shown in Figure 8b, participants express a low level of confidence in determining the location of the emergency shelter based solely on the guidance signs. Therefore, the displayed information and layout of the existing signs create uncertainty when selecting a route, potentially resulting in longer evacuation times and unwanted delays. In addition, the guidance signs are dispersed, indicating the need for more visual displays to enhance the continuity of these signs.

4. Discussion

The efficient evacuation of individuals from residential areas to a safe place, such as emergency shelters, poses a significant challenge for community management in the aftermath of a disaster. While there have been studies focusing on wayfinding behaviors in buildings, underground space, and other contexts, there is a gap in the literature when it comes to a systematic study of the physical behaviors and information perception of evacuees during the wayfinding process for emergency shelters. This study fills that gap by examining the wayfinding behaviors for emergency shelters and analyzing the attention distribution during the decision-making process reflected by eye metrics. A total of 17 participants were recruited for performing on-site experiments aimed at finding the nearest emergency shelter around the community.

Although directing signs are typically installed both inside and outside buildings, research has shown that fewer than 50% of individuals are able to perceive and comprehend emergency signs in presumed dangerous situations [4,10]. In our study, most participants noticed the guidance signs for emergency shelters, but we also observed that 17% of participants struggled to understand the content of these guidance signs due to poor alignment between the sign direction and complex intersections. Numerous studies have investigated the effective distance, direction compliance, design, location, and content of directing signs in buildings [28,33,35,36,37,38,39]. Our findings are consistent with previous research, which emphasizes the significant impact of the angle of interaction and arrow direction on the evacuees’ confidence level. Future studies should aim to improve the continuity and legibility of emergency shelter signs.

In building evacuation studies, researchers have incorporated the heterogeneity of evacuees’ walking velocity, route choices, following intention, and aggressiveness into simulation to achieve accurate results [41,42,43,44]. In this study, we have focused on understanding the complexity and heterogeneity of evacuees’ physical behavior and visual behavior during the task of finding an emergency shelter. By doing so, the behavioral principles derived from this study can also enhance the accuracy and reliability of simulation models. This contributes to the improvement of emergency preparedness and response strategies for outdoor evacuations.

With the advancements in wearable eye tracking technology, measuring the visual behavior of participants during the shelter-finding process has become convenient. However, it is important to acknowledge the limitations associated with the use of this device. Firstly, it requires participants to have good unaided vision, which can pose challenges during participant recruitment. Additionally, despite the calibration process being conducted to accurately locate participants’ eye positions before the experiment, there may still be biases and unstable outputs, leading to missing or drifting fixations for a few frames. Furthermore, the current device is limited in its ability to measure superficial interactions, and incorporating co-recordings of electroencephalography, electromyography, and other bio-signals would provide opportunities to uncover the perception and motivation behind the observed behavior [45].

To fully understand the processes of perception, decision-making, and action during crowd evacuation, more experiments and simulations need to be conducted. Specifically, the current study had a limited number of participants (17) involved in the emergency-shelter-finding task, making it challenging to generalize patterns of outdoor evacuation behavior. Therefore, in future studies, we plan to expand the scale of the participant population and conduct group experiments at the same time. We will also consider the influence of surrounding people and social networks in the experiment. The function of directing signs may differ between single-person experiments and multi-person experiments. Furthermore, this work only focused on one group of students from the campus. It is important to include other types of evacuees, such as residents of different ages in the community who are familiar with the region, in future studies. Currently, while this study analyzed typical eye metrics extracted from video to identify attention distribution, future research can incorporate image processing technique such as super resolution [46] and inpainting [47] to further investigate scanpaths and gaze heatmaps during crisis decision-making.

5. Conclusions

This study revealed the hesitation and retracing behavior of participants during the wayfinding process for emergency shelters. The attention distribution for the surrounding objects was analyzed using an eye-tracking device. A series of on-site experiment were conducted involving student participants, leading to the following conclusions:

(a)

The demographics of the participants and background factors have a critical impact on shelter-finding behaviors. There are notable differences in total hesitation duration (p = 0.045, mean difference = −13.89), detour duration (p = 0.034, mean difference = −105.50), and total experiment duration (p = 0.038, mean difference = −15.83) between different genders. Proactive participants exhibit a shorter extra route length ratio compared to the other types of participants (p = 0.007). Participants with a relatively rich experience of disasters and emergency shelters are more likely to find the nearest emergency shelter (p = 0.037).

(b)

The wayfinding behaviors are classified as a map-based style and a route-based style. The map-based group shows a higher average VAI (0.025) for the digital map, with a mean fixation duration of 244.17 ms, while the route-based group shows a higher VAI (0.053) for the road, with a mean fixation duration of 251.58 ms.

(c)

The individual visual behavior is obviously reflected in the total fixations of the participants, ranging from 38 to 639 throughout the entire experimental process. All participants spend approximately 30% of their total fixation counts on observing road conditions.

(d)

The findings also demonstrate the limitations of existing signs. For instance, the displayed information and layout of the signs create uncertainty during the route-selection process, indicating the need for optimization based on wayfinding behavior and attention distribution.

This study is the first step in addressing a complicated problem: the optimization of evacuation signs, shelter locations, and path designs, with consideration of people’s physical behaviors and cognition. The results obtained in this work have practical implications for the development of evacuation guiding systems and the planning of public drills, considering individual characteristics. The fundamental conclusions derived from this study can be valuable for government departments, social groups, and enterprises for improving urban resilience and ensuring city safety with respect to the utilization of emergency shelters.