Introduction

Safety is paramount in railroad development. The State Railway Administration of China’s Bulletin on Railway Safety Situation for 2019–2021 identifies the main causes of railroad accidents as foreign objects encroaching upon the tracks1. These objects include pedestrians illegally crossing the tracks, items entangled in the overhead lines such as color steel tiles, dust nets, and plastic debris from greenhouses along the railway, and obstacles on the tracks like livestock and tools forgotten by maintenance workers. Compared to other forms of transportation, trains frequently result in severe injuries during collisions due to their high speeds, lengthy stopping distances, and lack of directional agility. Accident statistics show that since 2000, there have been 25 significant railroad accidents globally, with 14 attributed to driver errors. This accounts for 56% of these incidents. Train operation is predominantly a visually demanding task, and trains continue to depend largely on manual operation. The driver’s ability to perceive hazards plays a crucial role in preventing accidents2,3,4.

Hazard perception, the crucial skill enabling drivers to identify, anticipate, and detect potential road hazards5. Various studies have demonstrated the effectiveness of this ability in forecasting a driver’s likelihood of being involved in accidents2,4. One factor contributing to a driver’s failure to visually scan for potential road hazards is the impact of workload6. Drivers subjected to hazard perception tests under different workload conditions7, exhibited a marked increase in reaction times to hazardous stimuli and a significant uptick in false alarms for non-hazardous stimuli when operating under high-load conditions. Konstantopoulos discovered that novice drivers experienced considerably higher workloads while driving at night or in rainy conditions, compared to experienced drivers under the same conditions8. Moreover, the perceived workload of train drivers may be influenced by the duration of the trip, and as the journey progresses, the workload is likely to increase. This increased workload resulted in a more limited horizontal visual scanning range and lengthened reaction times to potential hazards. Furthermore, Engström found that high workloads can diminish driving performance, evident through reduced speeds and more frequent variations in lane-keeping ability9. Drawing on the insights from the aforementioned studies, we hypothesize that the process of hazard perception in train drivers is likely influenced by their level of workload, which in turn impacts the effectiveness of hazard recognition. Train drivers’ workload may be influenced by the duration of the trip10. More importantly, the workload experienced by train drivers is closely associated with the speed of the train11. Travel on general-purpose railroads typically involves lower train operating speeds and monotonous line information, which results in relatively low workloads for train drivers. However, an inadequate workload can result in distraction and a diminished capacity to concentrate on essential tasks12. This reduced level of vigilance makes train drivers less responsive to perceived hazards in their environment. Consequently, they may be more prone to overlook potential risks or fail to react adequately to rapidly changing situations, which increases the risk of accidents13,14. In comparison to trainee drivers, experienced train drivers are better at balancing their visual focus between the external track environment and the monitoring interface15,16. They also maintain prolonged attention on the track ahead and have a lower likelihood of committing violations. On the other hand, train drivers who have been involved in a railroad accident exhibit lower levels of sustained attention17. These drivers are more prone to delayed hazard recognition if a similar situation arises again. Based on these findings, we propose Hypothesis 1: At general operating speeds, the stability of train drivers’ attention towards the track ahead is a crucial measure of their competency, which subsequently influences their performance in responding to hazards. In eye-movement research, the search time ratio, defined as the proportion of time a subject spends searching the forward visual field, is considered a vital indicator of an individual’s attentional stability18. This metric could serve as a significant predictor of a train driver’s ability to respond to hazards effectively.

In the context of China’s swiftly advancing railroads, the operation of high-speed railroads is marked by rapid train speeds and a substantial influx of line information. This scenario increases the workload for train drivers, imposing greater demands on their occupational competence, particularly in terms of their efficiency in processing information19. Prior research into dynamic visual search has revealed that an increase in speed substantially impacts observers’ eye movements and attentional control20,21, impairs participants’ performance in visual search tasks22. The train driver’s efficiency in monitoring crucial information both inside and outside the vehicle directly correlates with better decision-making performance23. Meanwhile, in high-speed driving scenarios, train drivers who sustain a longer gaze duration are more adept at recognizing track subsidence ahead of the moving train24. Consequently, Hypothesis 2 posits that a train driver’s ability to efficiently process track information at high speeds is crucial for their competence and influences their hazard response effectiveness. In eye movement research, mean gaze duration, defined as the average length of each gaze fixation within a unit of time, is considered a sensitive metric for assessing an individual’s efficiency in processing information25,26,27, can serve as a valid eye movement parameter for evaluating train drivers’ performance in responding to hazards28. On conventional railways, where train speeds are slower and track information is monotonous, it is increasingly important for locomotive operators to maintain sustained and stable attention. In contrast, at high speeds, the workload increases, demands on their competencies rise, and efficient information processing becomes more crucial for locomotive operators. Additionally, existing research has shown that the eye-tracking indicators representing sustained attention and information processing efficiency are distinct29. Therefore, theoretically, the eye-tracking indicators that affect locomotive operators’ hazard perception should differ across different speed environments. Identifying effective eye movement indices for train drivers operating at various train speeds is crucial. Such measures are key to aiding drivers in enhancing their attention control, and they offer scientific backing for boosting both the efficiency and safety of train drivers’ work. Theoretically, on conventional rail where the train speed is slower and the track information is monotonous, it is more important for train drivers to maintain sustained attention and vigilance. In contrast, at higher speeds, the workload increases, and the monitoring demands also increase, which places higher requirements on their competency. In this case, information processing efficiency becomes more critical, and the driver must react quickly once a hazard is detected. Existing research has shown that eye movement metrics representing sustained attention and information processing efficiency differ29. Therefore, eye movement indicators influencing hazard perception in train drivers should theoretically differ under different speed conditions. However, this hypothesis currently lacks empirical validation.

The alignment of the track a train traverses represents the most notable dynamic change in external environmental information. When operating on curved sections, a train driver’s workload escalates due to visual obstructions. This scenario necessitates heightened visual attention from the driver, who must invest more time in observing track conditions to gather sufficient information. Guo et al. found that the probability of collision between train and pedestrian increases significantly when the train runs on sharp curves30. At the same time, there is a left-leaning attention bias in hazard perception31, meaning that healthy individuals exhibit a slight leftward spatial attention asymmetry32. In a related study in the field of road traffic, Benedetto et al. conducted lane change test experiments in France, where lane changes were guided by information provided by two identical road signs displayed simultaneously on the left and right side of the road31. The study revealed that participants predominantly focused their attention on the left side of the sign. Similarly, a general bias towards the left side of space in drivers’ visual attention was observed in an experiment conducted in China33. In the realm of railroad transportation, Ma et al. conducted experiments to assess the hazard sensitivity of train drivers by comparing those with and without experience of foreign object encroachment28. By utilizing videos depicting scenarios of foreign object encroachment on railroads, generated through a driving simulator, they discovered that train drivers exhibited heightened hazard sensitivity to foreign objects on left curves as compared to right curves. Based on these findings, we propose Hypothesis 3: train drivers are likely to detect foreign objects more quickly on left turns compared to right turns, and this left-leaning attentional bias in train drivers may vary depending on the amount of information presented at different train operating speeds. As train speed increases, the left-leaning attentional bias of train drivers might become more pronounced. Recognizing this bias is crucial for optimizing the arrangement of line and signal signs in the field of railroad transportation. For example, it is relevant to investigate whether there is a difference in the effectiveness of traffic lights placed on the left and right sides of a train. Additionally, during train operation, the direction from which danger approaches is random, which makes considering left-leaning attentional bias practically significant in the safety training of drivers.

As previously discussed, an efficient visual search pattern is essential for train drivers to swiftly and accurately identify potential hazards. The capacity of drivers to quickly recognize foreign objects on the track during operation is a critical component of driving safety. Consequently, this study aims to uncover the differences in eye-movement predictors and the left-leaning attention bias on train drivers’ hazard response performance under different speed conditions. This will be achieved through experimental methods and the analysis of eye-movement data. The findings of this study are intended to provide a theoretical foundation and reference for the creation of a specialized hazard perception training program for train drivers, aimed at further enhancing the safety of train operations.

Method

Participants

In this study conducted at the Driver Training Base of the Lanzhou West Engine Section in Lanzhou City, Gansu Province, a cohort of 60 train drivers was selected to participate as subjects. The subjects, all of whom were male, were randomly assigned into two distinct groups: the general speed group (comprising 30 drivers) and the high-speed group (also comprising 30 drivers). We calculated the required sample size using G*Power, targeting a medium effect size (f = 0.25), 90% power, and an alpha of 0.05. The design involved two groups and nine repeated measures. Assuming a correlation of 0.5 among measures, we determined that a minimum of 16 participants was necessary.

All subjects were in good physical health and had been working on the front line of the railroad for the past month and had actually followed the train. All subjects had a good work record without any serious violations and had not participated in similar experiments. The subjects had 20/20 vision or corrected vision, had not consumed alcohol or any medication that might affect their cognition or mood in the week prior to the test, and possessed proficiency in using the mouse to respond. All subjects signed an informed consent form prior to participating in the experiment, indicated their willingness to actively participate, and were in good mental health.

The research procedure and data collection complied with the American Psychological Association ethical standards and the 1964 Helsinki Declaration with its later amendments. This study was approved by the Research Ethics Committee of Southwest Jiaotong University.

Experimental material

The simulation platform used in this study featured a highly realistic train driver perspective of foreign object encroachment, developed by Southwest Jiaotong University. This simulator, equipped with a 6-degree-of-freedom motion system, boasts a single-channel large-screen forward view offering an 160° horizontal field of view. Its effectiveness has undergone thorough vetting through systematic evaluations, which confirm its aptitude for meeting our experimental criteria (see Fig. 1a). The video materials utilized were of high-definition quality, with a resolution of 1920 × 1080 pixels in CVI format. These videos underwent meticulous refinement using After Effects, with each test video centered on a singular foreign object encroachment hazard. This included three main types of scenarios: static foreign objects positioned centrally on the track, and scenarios featuring left curves, straight paths, and right curves, all with equal curve radii (refer to Fig. 1b, c, and d). The duration of each test video ranged from 0 to 60 s, with the appearance time of the foreign object randomized. To subtly veil the experiment’s purpose, some videos without foreign object incidents were included in the formal experiment, constituting 50% of the total videos. Additionally, the presentation of experimental materials was randomized to eliminate any biases due to practice or fatigue effects.

Fig. 1
figure 1

Driving simulator and three kinds of foreign body invasion danger situations. Image source: Generated using the High-Speed Railway Driving Behavior Safety Simulation Platform developed by Southwest Jiaotong University.

Full size image

There are many types of foreign objects that encroach on railway limits. Regarding the selection of foreign object types, this study refers to the “Railway Safety Situation Announcements” published by the China National Railway Administration from 2019 to 20211. These announcements indicate that pedestrians illegally crossing the tracks, foreign bodies in the overhead contact line (such as color steel tiles along the line, dust nets, and remnants of greenhouse plastic), and track obstacles (such as livestock and tools left by maintenance workers) are major causes of railway accidents. Additionally, through interviews with locomotive crews, considering factors such as the frequency of foreign object encroachments during regular driving, the relationship between foreign object danger and accidents, and the number of foreign object dangers occurring simultaneously, this study selects the following three main types of foreign object encroachments: Pedestrians illegally crossing the tracks, Foreign body in overhead contact line, and Track legacy tool (Table 1).

Table 1 Frequency and proportion of foreign body intrusion risk situations in formal test videos.
Full size table

In the video materials tailored for the general speed group, the depicted train consistently travels at a steady pace of 120 km/h. Conversely, for the high-speed group, the train in the videos maintains a constant velocity of 240 km/h. Criteria for Grouping refers to the “Technical Management Regulations for Railways—High-Speed Railway Section” published by the China Railway Corporation34. Although there are differences in the operations of high-speed and conventional train drivers during actual train operations, the task in this study involves a simple perceptual task. During the initial lookout phase of train driving, the emergency response of both groups of drivers is to apply the brakes. Each group is represented by a set of 30 videos showcasing various scenarios of foreign object encroachment on the tracks. Of these, 3 videos per group were designated for practice tests, while the remaining 27 were reserved for the formal experiment. Table 2 details these 27 formal test videos, including types and descriptions of the foreign objects encountered. Notably, the video materials for both the general speed and high-speed groups feature identical kilometer markers, with the only variable being the speed of the train.

Table 2 Basic demographic information and comparative analysis of subjects (N = 60).
Full size table

Experimental device

The Tobii X120 eye-tracker, engineered by Sweden’s Tobii, is an advanced, standalone eye-tracking device offering a 120 Hz sampling rate, with a precision of 0.3° and an accuracy of 0.5°. Designed for versatility, the Tobii X120 is compatible with a broad range of populations and accommodates variations in ethnicity, age, and eyewear, including both glasses and contact lenses. This eye-tracker allows subjects to move their heads freely during experiments, eliminating the need for restrictive headgear, headrests, or external cameras. Consequently, participants can undergo the eye-tracking test in a fully natural state and setting. This enhances both comfort and the authenticity of the results. This device was used to record the participants’ eye movements during the experiment.

For the experiment, a 27-inch computer display was utilized, featuring a screen resolution of 1920 × 1080 pixels. The eye tracker is placed below the screen. During the actual experiment, participants were asked to sit approximately 60 cm in front of the eye tracker. The test videos were presented and data acquisition was managed through Tobii Studio 3.0 software. This advanced software is designed to automatically record the subjects’ reaction times and eye movement indices. This streamlines the data collection process and enhances the accuracy of the results.

Experimental design

A three-factor mixed experimental design of 2 (train speed: general speed group, high speed group) × 3 (track type: left curve, straight track, right curve) × 3 (foreign object type: pedestrians illegally on the track, tools left on the track, and foreign objects on the contact network) with train speed as the between-groups variable, track type and foreign object type as the within-groups variables, with the dependent variables being the proximity to the collision distance and the eye movement index.

Evaluation index

Approach collision distance indicator

Numerous studies have pointed out that the definition of a driver’s hazard perceptual response must be strictly based on a defined cutoff criterion35,36,37, as shown in Fig. 2. In the field of road traffic, many studies use Time to Collision (TTC) to assess the risk of collision. However, in rail transport, TTC may not be a clear indicator with specific safety guidance implications. For vehicle collisions, the time interval from the onset of danger to the collision is often very brief. TTC can easily convey the urgency of the approaching hazard, but on long tracks, it takes a considerable amount of time from when a train driver spots a foreign object hazard from afar and takes emergency braking actions to when a collision actually occurs. During this process, the train’s speed decreases rapidly, and the braking time varies with the type of train, load, and speed in different actual driving scenarios. Hence, the urgency represented by TTC can only be a relative value. This makes it difficult for safety managers to understand and confirm its urgent significance. The present study draws on the definition of TTC in the field of road traffic and proposes the distance to near-collision (Distance to near-collision, DTNC, unit: m): the distance from the collision point to the location where the train driver recognizes the danger and responds with a key press. Compared to Time to Collision (TTC), the distance from when the driver detects a hazard to when a collision occurs (DTNC) is fixed, offering a clearer indication of imminent danger.

Fig. 2
figure 2

Response window definition.

Full size image

In this study, the speed is constant, allowing for the calculation of the distance between the driver’s braking response and the foreign object. This makes this indicator more referential for accident prevention efforts. In the field of traffic safety, numerous studies have indeed utilized reaction distance as a metric for assessing hazard perception capabilities38,39. Moreover, our experiment used video materials. For static images, the time of hazard appearance can be set based on image presentation time. However, in video materials, since the scenes are dynamically changing, hazards do not appear from the first frame. In reality, it is technically difficult to determine the exact moment a hazard appears, and this is even more challenging in actual train driving scenarios. For instance, different drivers have varying visual acuity, sensitivity to hazards, judgment criteria, and perceptual thresholds. The exact moment a hazard stimulus becomes barely visible is highly subjective, making it difficult to pinpoint the onset of the hazard. Therefore, hazard reaction latency can only be a relative measure. Instead, we used the distance between the participant’s key-press braking response and the collision point of the foreign object with the train. This distance is an objective, absolute value that can be precisely calculated. Therefore, this study adopts the DTNC metric. The longer the DTNC is, the earlier the subject recognizes the foreign object, and the longer the time left for braking. If no braking response is initiated before a collision with a foreign object, this data is marked as 0 in the calculations, indicating the weakest hazard perception ability. If the subject’s key press time point falls outside the response window, it is considered an invalid response and is not counted in the statistical analysis. Because widespread reactions only began after the starting point of the response window. Responses occurring before this point were most likely due to operational errors, accidental triggers, or inattentive reactions. Additionally, we examined all instances of premature responses from participants, and the data showed that these reactions occurred at time points far removed from the start of the response window. This indicates that they were not made in anticipation of the hazard. Response windows and response rates for different hazardous scenarios are shown in the Appendix.

Eye tracking index

Train drivers need to keep a constant lookout for the road ahead. In order to explore the visual attention characteristics of the train driver during the hazardous search phase, the time window for analyzing the eye movement data started 1 s after the beginning of the video playback and continued until one frame before the appearance of the foreign object. Then, we analyzed the train driver’s eye movement heat maps (Fig. 3 d, e, f) based on the type of track where the foreign object was located, which was divided into three scenarios: a left curve scenario, a straight track scenario, and a right curve scenario (Fig. 3 a, b, c); the region of interest covered the center area of the forward visual scan (Fig. 3 g, h, i), and the entire scene area was 1920 × 1080 pixels, and each region of interest was of equal area, both of which were 300 × 300 pixels.

Fig. 3
figure 3

Scene, hotspot map and area of interest under different track types. Image source: Generated using the High-Speed Railway Driving Behavior Safety Simulation Platform developed by Southwest Jiaotong University.

Full size image

Total Visit Duration Percentage (TP, unit: %): This metric represents the percentage of time spent searching the forward view in the video segment before the appearance of a foreign object. The denominator of the search time ratio is the total viewing time of the video segment, which includes the duration of all field of view fixations, while the numerator is the time specifically spent by participants searching the forward field of view before the appearance of the foreign object. A higher percentage indicates that the locomotive crew spent more time searching the forward view.

Fixation Duration (FD, unit: s): This metric represents the average duration of gaze at all fixation points in the forward view in the video segment before the appearance of a foreign object. The numerator is the total fixation time on the forward field of view, and the denominator is the number of fixation points. A longer average fixation duration indicates that the locomotive crew spent more time processing information in that area.

Experimental procedures

To ensure that participants were in a clear and alert mental state for the hazardous perception experiment and to minimize any physiological interferences with the experimental data, the experiment was scheduled to begin at 9 a.m. each day. Initially, participants filled out an informed consent form and a basic demographic information survey. Next, the experimenter explained the experiment instructions in detail, including the guidance for the experimental tasks. Participants were required to continuously monitor track conditions and press a button upon detecting a foreign object. It was important to note that participants should not click the left mouse button repeatedly during a single video; the video would continue playing even after a click, and each video would automatically proceed to the next one upon completion. Before the official experiment commenced, participants engaged in a practice test to ensure they fully understood the experimental task instructions. The experimenter used a 5-point calibration method to calibrate the participants’ eyes. This ensured the accuracy of the eye movement tracking. After successful calibration, the formal experiment began. To eliminate any practice and fatigue effects, the experimental materials for each participant were presented in a random order. The experimental scene and equipment setup are illustrated in Fig. 4. Throughout the entire experimental process, the experimenter closely monitored the participants to ensure the experimental environment was free from external disturbances. The overall experimental workflow is depicted in Fig. 5. The experiment lasted approximately 30 min.

Fig. 4
figure 4

Experimental scenes and equipment of subjects. Image source: Photographs taken during the experiment.

Full size image
Fig. 5
figure 5

Flow chart of the experiment.

Full size image

Data analysis

When analyzing demographic factors, proximity to collision distances, and eye-tracking indicators of locomotive operators, the varying scales and units of these metrics may introduce errors into the data analysis results. To eliminate these errors, we have employed the Z-score normalization method, which involves subtracting the mean from each data value and then dividing by the standard deviation. This process normalizes the distribution of the dataset to a mean of 0 and a standard deviation of 1.

Firstly, the effects of vehicle speed group, track type and foreign object type on the risk response performance of train drivers were investigated. The repeated measurement ANOVA was conducted with the vehicle speed group as the inter-group variable, track type and foreign object type as the intra-group variable, and proximity collision distance as the dependent variable. Then, the visual attention characteristics of different track types on train drivers in the danger search stage, where they focus on detecting potential hazards along the track, were analyzed. The speed group was taken as the inter-group variable, the track type was taken as the intra-group variable, and the ratio of search time and average fixation time were taken as the dependent variable. Finally, the relationship between eye movement index and hazard response performance was established by regression equation under different speed groups, controlling the track type. It was concluded that the search time ratio at normal speed could positively predict the hazard response performance of train drivers, while the average fixation time at high speed could negatively predict the hazard response performance of train drivers.

Results

In this study, IBM SPSS 27.0 was used as a statistical analysis tool, and the repeated measures ANOVA and regression analysis to statistically and analytically analyze the data.

Internal consistency reliability

Taking DTNC as the index, the internal consistency coefficient of the 27 test videos is calculated to be 0.897. Subcategorized according to the track type and foreign object type, the internal consistency coefficients of the left curve, straight track, and right curve are 0.829, 0.817, and 0.678; and those of the pedestrian illegally boarding the roadway, the foreign object on the contact network, and the tools left on the track are 0.815, 0.768, and 0.734. Generally speaking, the internal consistency coefficient above 0.60 is reasonable, which indicates that the reliability test result of this test is good.

Demographic comparison of driver groups

Comprehensive demographic data and a comparative analysis of participants across these two groups are methodically presented in Table 1. A thorough examination revealed no statistically significant disparities between drivers of the general speed group and those in the high-speed group in terms of crucial demographic parameters such as age, length of service, educational background, and history of foreign body intrusion accidents. This finding underscores a remarkable demographic parity and absence of significant imbalance between the two groups with respect to these pivotal factors.

Approach collision distance

To investigate the relationship between DTNC of train drivers of different speed groups to different foreign objects on different track types. Repeated-measures ANOVA revealed significant main effects for track type (F (2, 57) = 123.036, p < 0.001, ŋp2 = 0.812), foreign object type (F (2, 57) = 542.453, p < 0.001, ŋp2 = 0.903), and vehicle speed group (F (1, 58) = 57.864, p < 0.001, ŋp2 = 0.499). The interaction between track type and vehicle speed group was significant (F (2, 57) = 25.203, p < 0.001, ŋp2 = 0.469). The interaction between foreign object type and vehicle speed group was significant (F (2, 57) = 10.602, p < 0.001, ŋp2 = 0.155). The interaction between track type and foreign object type was significant (F (4, 55) = 112.452, p < 0.001, ŋp2 = 0.891). The three-way interaction between track type, foreign object type, and vehicle speed group was significant (F (4, 55) = 20.271, p < 0.001, ŋp2 = 0.596). Simple effects analyses indicated (see Figs. 6 and 7):

Fig. 6
figure 6

Simple effect analysis of the interaction of track type and foreign body type on DTNC in the normal speed group.

Full size image
Fig. 7
figure 7

Simple effect analysis of interaction between track type and foreign body type on DTNC in high-speed group.

Full size image

For the general speed group, there was a significant difference in the DTNC for the pedestrian illegal on-road encroachment scenario for the three track conditions (F (2, 28) = 190.211, p < 0.001, ŋp2 = 0.931), with the longest DTNC for the straight roadway (M = 325.742, SD = 14.516), followed by the left curve (M = 234.872, SD = 4.621), and the shortest DTNC for the right curve had the shortest DTNC (M = 131.439, SD = 4.619); there was a significant difference between the DTNCs for the foreign object encroachment cases on the contact network (F (2, 28) = 46.553, p < 0.001, ŋp2 = 0.616), with the longest DTNCs for the straight course (M = 287.838, SD = 8.707), the left curve (M = 195.397, SD = 4.548) and no significant difference in DTNC for the left (M = 195.227, SD = 10.234) and right (M = 195.227, SD = 10.234) curves; there was no significant difference in DTNC for the legacy tool encroachment condition (F (2, 28) = 1.507, p = 0.23 > 0.05, ŋp2 = 0.049).

In the high-speed group, there is a significant difference in DTNC for the pedestrian illegal on-road encroachment case (F (2, 28) = 115.071, p < 0.001, ŋp2 = 0.892) under the three track conditions, with the DTNCs for the straight track (M = 193.364, SD = 15.975) and the left curve (M = 199.634, SD = 6.974) being larger than that for the right curve (M = 107.088, SD = 3.127), with no significant difference in DTNC between straight and left curves; no significant difference in DTNC for the foreign object encroachment case on the contact network (F (2, 28) = 1.293, p > 0.05, ŋp2 = 0.043); and a significant difference in DTNC for the legacy tool encroachment case (F (2, 28) = 20.155, p < 0.001, ŋp2 = 0.590), with the straightaway having the longest DTNC (M = 76.649, SD = 5.958), followed by the left-hand curve (M = 55.664, SD = 3.472), and the right-hand curve having the shortest DTNC (M = 45.789, SD = 2.650).

Eye movement

In order to explore the visual attention characteristics of train drivers in the dangerous search stage, the eye movement indicators selected in this study were before the appearance of foreign bodies, independent of specific types of foreign bodies.

Our study confirms that average fixation, reaction time, and overall number of fixations can all serve as effective measures of information processing efficiency. The correlation analysis revealed that these three indicators are interrelated: average fixation duration shows a positive correlation with reaction time (r = 0.37, p < 0.01), a negative correlation with number of fixations (r =  − 0.59, p < 0.001), while reaction time demonstrates a negative correlation with number of fixations (r =  − 0.86, p < 0.001). To avoid potential multicollinearity issues arising from the high correlations among these independent variables, we have focused specifically on average fixation duration as the primary independent variable in our subsequent analyses.

Search time ratio (TP)

As shown in Fig. 8, an ANOVA was conducted to analyze the TP of the two groups of train drivers on different track types, and the results showed that the main effect of track type was significant (F (2, 57) = 12.547, p < 0.001, ŋp2 = 0.178). Post hoc tests showed that train drivers had significantly smaller TP on straights than on curves, and significantly larger TP on left curves than on right curves. The main effect of speed group was significant (F (1, 58) = 6.427, p = 0.014, ŋp2 = 0.100). TP was significantly higher in the high speed group than in the general speed group. The interaction between track type and speed group was not significant (F (2, 57) = 0.032, p = 0.969, ŋp2 = 0.001).

Fig. 8
figure 8

Mean and standard deviation of search time ratio of train drivers under different track types.

Full size image

Average fixation time (FD)

As shown in Fig. 9, an ANOVA was conducted to analyze the FD of the two groups of train drivers on different track types, and the results showed that the main effect of track type was significant (F (2, 57) = 21.268, p < 0.001, ŋp2 = 0.427). Post hoc tests showed that train drivers had significantly greater FD on straights than on curves, and significantly greater FD on left curves than on right curves. The main effect of speed group was significant (F (1, 58) = 7.961, p = 0.007, ŋp2 = 0.121). FD was significantly higher in the high speed group than in the general speed group. The interaction between track type and speed group was not significant (F (2, 57) = 0.534, p = 0.589, ŋp2 = 0.018).

Fig. 9
figure 9

The mean value and standard deviation of the average gaze time of train drivers under different track types.

Full size image

Prediction function of eye movement index

Under different track types, eye movement index was used as the predictor variable and DTNC was used as the result variable for hierarchical regression analysis. Using the ENTER method, the track type variable is introduced into the first layer, and the search time ratio and average fixation time are introduced into the second layer. The results are shown in Table 3.

Table 3 Regression table of eye movement index to DTNC.
Full size table

Through hierarchical regression analysis, it is found that after controlling the track type, the search time ratio can positively predict the DTNC of train drivers and the average fixation time at high speed can negatively predict the DTNC of train drivers.

Discussion

Problems and analysis of risk response performance of train drivers under different speed conditions

Our results found a three-way interaction between speed factor, track type, and foreign object type on train drivers’ hazard response performance. In conducting hazard detection tasks at different speeds, train drivers exhibited shorter Detection Time to Navigational Challenge (DTNC) on curves compared to straight tracks, highlighting a marked degraded curve recognition. We hypothesize that this phenomenon might be attributed to two factors. First, curves inherently obstruct the line of sight, requiring train drivers to anticipate ahead. Second, as indicated in studies like Huo40, navigating curves increases the cognitive load on drivers. The difference in DTNC for the curves compared to the straights was more pronounced in the foreign object context where sensory salience was higher. This further indicates that the degraded curve recognition was more pronounced in this context. Sensory saliency refers to certain perceptual properties of the stimulus itself that make the observer’s attention involuntarily drawn to it41. These perceptual attributes encompass aspects such as stimulus intensity, motion changes, novelty, and contrast relations. Previous research has demonstrated that the perceptual saliency of a stimulus influences an observer’s visual search efficiency. For instance, Huang & Chiu found that the response time for identifying a circular icon was considerably shorter compared to a triangular one42, and icons with a border line width of 3 pixels were identified more quickly than those with 1 or 2 pixel widths. Similarly, Shive et al. reported that selecting low saliency colors led to slower search times for items on displays with high clutter, as opposed to those with low clutter43. In this study, we established three categories of foreign object encroachment types. Track pedestrians were found to possess the highest perceptual convexity, attributed to their prominent three-dimensional standing features. Contact network foreign objects ranked second in terms of perceptual convexity, while track leftover tools exhibited the lowest perceptual convexity, primarily due to their less noticeable three-dimensional features as they lay flat on the railroad tracks. With the decrease in the perceptual saliency of these foreign objects, our attention naturally shifts towards more salient targets or areas. Consequently, we postulate that the degraded curve recognition is more pronounced for foreign objects that exhibit higher perceptual convexity.

And we also found that the DTNC of the straightaway was shorter as the speed increased, indicating a decrease in the straightaway recognition advantage. We hypothesize that this may be due to the fact that trains run faster and have increased workloads under high-speed conditions, and at the same time, the same reaction time results in a shorter DTNC. As a result, the train driver’s response performance is reduced at high speeds in the straightaway. Simultaneously, we found that the degraded curve recognition becomes more accentuated with increasing speed. This could be attributed to the heightened demands placed on the train driver’s perception and cognition under high-speed conditions. As trains operate at faster speeds, the workload on drivers intensifies. This makes train drivers focus more intently on the track state directly within their central field of view and reduces their scanning of peripheral track information to avoid overlooking critical details. This focused approach amplifies the degraded curve recognition at higher speeds by making it more challenging to identify foreign objects on curves. As a consequence, train drivers exhibit diminished response performance at high speeds, with the degraded curve recognition being particularly pronounced on curves. This phenomenon aligns with findings from Guo et al.44, who noted that as train speed increases, train drivers’ attentional focus progressively narrows to the track portion in the center of their visual field, while the rate of attention to the surrounding environment decreases logarithmically. This shift in visual attention pattern may represent an adaptive attentional strategy employed by train drivers to cope with the demands of a high-speed environment.

In summary, there are differences in the hazard detection performance of train drivers under different speed conditions, which may be due to different speeds and job task requirements (different visual modalities), and therefore there may be differences in the predictors of train driver hazard performance under different speed conditions.

Differences in eye movement prediction indexes of risk response performance of train drivers under different speed conditions

The aforementioned analysis revealed that train drivers’ hazard response deficits manifest differently under different speed conditions. Given that eye movement indices are indicative of the visual processing and attention allocation processes in train drivers, the question arises: Can these eye movement indices predict train drivers’ hazard response performance under different speed conditions? Hierarchical regression analysis revealed that, within the general-speed group, the ratio of search time positively predicted train drivers’ hazard response performance, whereas for the high-speed group, average gaze time had a negative predictive relationship with their hazard response performance. This aligns with the findings of Suzuki et al.24, who investigated the visual search behaviors of train drivers in identifying abnormal events (such as track subsidence ahead of the train’s path). Their study concluded that in low-speed scenarios (approximately 15 km/h), train drivers with a broader horizontal field of view more effectively recognized track subsidence. Conversely, in high-speed scenarios (~ 90 km/h), those maintaining longer gaze durations were more adept at recognizing subsidence. These insights are consistent with our overall regression results. They underscore the differing eye movement metrics that predict train drivers’ hazard response performance in high-speed versus low-speed driving conditions.

Firstly, our study determined that the search time ratios positively predict train drivers’ hazard response performance under generalized speed conditions. The search time ratio is an indicator of the stability of a train driver’s visual attention towards the forward field of view. This metric has been previously demonstrated to characterize train drivers’ visual search behavior effectively, particularly in relation to indoor and outdoor equipment and control panels29,45,46. Simultaneously, this indicator has been found to be highly predictive of accidents. Train drivers exhibiting a high level of sustained attention are less likely to be involved in accidents47. However, the underlying reasons why this indicator is predictive of accidents have not been extensively explored in previous studies. We hypothesize that this may be due to the metric’s characterization of the visual attentional stability of train drivers focusing forward. In tasks involving railroad track monitoring, train drivers who can maintain their attention for extended periods tend to demonstrate better decision-making performance48. A similar observation was made in our study, where under generalized speed conditions, a larger search time ratio correlates with better hazard response performance. This metric reflects the train drivers’ ability to maintain their attention over prolonged periods, which enhances their hazard perception. This supports Hypothesis 1. Since hazard perception forms the foundation of driving decisions, this improvement in perception directly contributes to better driving performance. This correlation supports the interpretation of the search time ratio as a predictive metric for accident likelihood. It highlights its significance in understanding and improving train driving safety. In addition, search time ratios did not predict hazard response performance in high-speed train drivers, may be related to the characteristics of general-speed railroad driving. Regular-speed railroads operate at slower speeds, and train drivers are more prone to distraction and reduced alertness during prolonged driving49,50. In this environment, train drivers’ visual attentional stability to the track ahead when recognizing foreign object encroachment situations is a key indicator of their competence. Therefore, this study identifies a validated eye movement metric that can be used to measure train drivers’ attentional stabilization ability during general-purpose railroad travel.

Secondly, our study found that under high-speed conditions, mean gaze time negatively predicts train drivers’ hazard response performance. Mean gaze time serves as an indicator of how efficiently a train driver processes information from the forward field of view. This concept is corroborated by previous research; for instance, Du et al. observed that novice train drivers exhibited longer mean gaze times compared to their experienced counterparts in information acquisition25, which hindered their hazard recognition capabilities. This finding suggests that experience influences the visual behavior of train drivers. It gradually transitions their information acquisition strategy from prolonged to short-term gazing. Such a shift in visual strategy can enhance driving performance, particularly in high-speed scenarios. Madleňák et al. found that the train driver’s average gaze time for recognizing ground signals was the longest when the train was traveling between zones26, while the train driver’s average gaze time for the panels with control functions of the train was the longest when the train was traveling in stations. This is due to the fact that the information on ground signals and panels with control functions is more complex and important and requires more processing time, hence their longest average gaze times. The present study arrives at a similar conclusion, indicating that a shorter average gaze time correlates with better hazard response performance under high-speed conditions. This finding suggests that the metric is an effective characterization of the information processing efficiency of train drivers. In essence, faster information processing by train drivers leads to superior hazard recognition performance, supported Hypothesis 2. Additionally, it was observed that average gaze time did not predict hazard response performance for general-speed train drivers, which we hypothesize could be related to the unique characteristics of high-speed railroad operation. High-speed railroads operate at elevated velocities and impose more perceptual and cognitive demands on train drivers. This makes them more vulnerable to operational performance degradation11. Compared to their general-speed counterparts, high-speed train drivers tend to exhibit enhanced performance in emergency situations, a critical aspect for the safety of high-speed railroads51. In such a high-speed context, the ability of train drivers to efficiently process information about the track ahead becomes a crucial indicator of their competence. Therefore, this study identifies an effective eye-tracking metric that can be employed to assess train drivers’ information processing capabilities during high-speed railroad travel.

The results of this study reveal variations in the predictive capabilities of train drivers’ eye movement metrics at different travel speeds. Future research could utilize these insights into eye movement characteristics under varied travel conditions to develop more precise training and attention regulation strategies for train drivers. Secondly, future research could more comprehensively examine the factors that may influence train drivers’ driving safety, in order to gain a deeper understanding of hazard perception performance. Moreover, including physiological signals such as Electrocardiogram (ECG) and Electroencephalogram (EEG) could provide a more comprehensive understanding of the cognitive states of train drivers. This would deepen our grasp of the mechanisms behind their hazard response performance. To prevent train drivers from overlooking potential dangers on the track, the implementation of a driver state detection system in the cabin could be considered. Drawing inspiration from Driver Monitoring Systems (DMS) used in automotive driving and adapting it to the unique aspects of railroad operation, such a system would continuously monitor and alert the state of train drivers52. This approach not only promises to enhance safety but also aids in understanding and improving the operational performance of train drivers under different conditions. However, although the eye-tracking metric of average gaze time is widely used in driving research to measure participants’ information processing level, gaze duration does not necessarily equate to the depth or extent of information processing. Gaze time may also be associated with participants’ attention allocation, visual habits, or other non-cognitive factors. Therefore, researchers in other fields should exercise caution when referencing these findings.

Left-leaning attention bias effect of train drivers

In our study, it was observed that train drivers responded more rapidly to pedestrian illegal encroachment situations on left curves compared to right curves during a general hazard detection task. The results supported Hypothesis 3. Supporting this hypothesis, previous research has demonstrated that drivers, irrespective of their seating position, exhibit a left-leaning attentional bias31,33. Consequently, we propose that train drivers, akin to other drivers, may unconsciously or naturally tend to focus more attention on the left side, indicating a similar left-leaning attentional bias28. With an increase in speed, our study noted an intensified leftward attentional bias in train drivers when recognizing ground-based foreign object encroachment situations. We speculate that this may be due to the low train running speed under the normal speed, which has a corrective effect on foreign bodies with poor sensation prominence. Additionally, the study revealed that under left curves, compared to right curves, train drivers demonstrated higher search time ratios and extended mean gaze durations. Our research revealed that train drivers displayed increased search time ratios and extended mean gaze durations under left curves as opposed to right curves. This tendency aligns with findings in the study of automobile drivers’ search behaviors. For instance, Benedetto et al. investigated the capacity of automobile drivers to recognize road signs31. They uncovered that 90% of eye movements were directed toward the left sign, while only 10% were directed toward the right. These eye movement metrics in train drivers reinforce the notion of a left-leaning attentional bias. They suggest a recognition advantage toward the left side. Given the train driver’s left-leaning attention bias, it is advisable to strategically position crucial wayfinding and signaling signs predominantly on the left side of the tracks. This placement ensures that drivers can access vital information promptly. However, this bias might lead to a degraded curve recognition on right curves. Consequently, enhanced focus should be placed on improving train drivers’ skills in identifying foreign objects on right curves during training and coaching sessions. Simultaneously, the foreign object detection systems, including sensors and cameras, should be oriented to monitor the right side of the tracks more vigilantly. This adjustment is crucial because train drivers are more prone to overlook potential hazards on this side, potentially leading to delayed recognition of such dangers.

The absence of a left-leaning attention bias in train drivers when recognizing foreign objects on the contact network is intriguing. This phenomenon might be inherently linked to the foreign object’s position. Given that the contact network is situated above the locomotive and the foreign object is suspended from it, train drivers may not exhibit a left-leaning bias when encountering airborne objects. The reason for a left-leaning attention bias towards ground-based foreign objects, but not for those airborne, remains unclear. Further validation is indeed necessary to confirm this observation. We need more comprehensive observations and analyses of real-world situations to ensure the accuracy of this speculation. Continuing in-depth research will enhance our understanding of how train drivers allocate attention and make decisions in response to foreign objects on the contact network. Such insights will be invaluable for developing more effective training programs and safety measures in the future.

Although this study has certain theoretical and practical significance, it still has several limitations. First, all the data in this study were obtained through simulation experiments and were not compared with representative real-world experimental data, so the external validity of the model needs further improvement. In future research, investigators could collaborate with railway operators to obtain real-world eye-tracking data (e.g., recordings from onboard eye trackers) during actual train operations involving foreign object intrusions. By conducting correlation analyses with simulation data, the validity of the simulation model could be verified, thereby enhancing its external validity. Second, the simulation experiments involved numerous participant-related variables. Future researchers should implement stricter control over extraneous variables, such as individual differences in visual acuity, while also appropriately expanding the sample size to improve the reliability of the findings and statistical power. Additionally, since data points where no braking response was initiated before collision with a foreign object are marked as 0 in our calculations (indicating the weakest hazard perception ability), future studies adopting this methodology should pay particular attention to the proportion of zeros in the dataset. If zero values constitute a substantial portion, the potential impact of zero-inflation on data analysis should be statistically addressed.

Conclusion

This study, leveraging videos of foreign object intrusions, gathered hazard response performance and eye movement data from train drivers undertaking hazard detection tasks at different speeds. This approach helped uncover differences in the eye movement predictors and the left-leaning attention bias on train drivers’ hazard response performance under different speed conditions. The empirical analysis highlighted distinct predictors for hazardous reaction performance contingent on speed. At general speeds, with slower train operation, train drivers need heightened attentional stability. We used the search time ratio to measure the stability of train drivers’ visual attention toward the forward field of view. The results showed that this metric positively predicts hazard response performance in the general-speed category.

Conversely, at high speeds, the efficiency of information processing becomes crucial. The mean gaze time, indicative of how quickly train drivers process information in the forward field of view, negatively predicts hazard response performance in the high-speed group. Thus, this study identifies two significant eye movement metrics for assessing train drivers’ hazard perception abilities.

Moreover, the research observed a left-leaning attentional bias in train drivers when recognizing ground-based foreign object intrusions. These findings provide a valuable theoretical foundation for improving train drivers’ visual attention capabilities. They also support the optimization of track and signal markings and the design of effective foreign object intrusion detection systems.

Future research avenues include deeper exploration of factors affecting train drivers’ recognition of foreign objects and the development of hazard perception detection tests and training tools. Although this experiment used simulated foreign object intrusion scenarios with relatively straightforward tasks, real-world driving is more complex. Hence, future studies could enhance ecological validity through more sophisticated driving simulation experiments.