Enhancing cyclist safety in cyclist-vehicle interactions through early hazard notifications: a comparison of bi-modal cues at head level

Hazard detection

Hazards can be detected by various means, such as on-bike sensors, e.g. radar (Engbers et al., 2018; Garmin Ltd, 2024; Kalaiselvan, 2021; Lindström et al., 2019) or by acoustic ranging (Jin et al., 2021). Another option would be to use cloud services linked to an accident database to indicate accident black spots (Trösterer et al., 2022). Further, future technologies enabling C-ITS (Hernandez-Jayo et al., 2015; Scholliers et al., 2014; Scholliers et al., 2017; Silla et al., 2017) will create opportunities to provide safety-related information to road users. Data obtained could be used, for example, to support cyclists during crossing decisions at intersections (Matviienko et al., 2019; Oczko et al., 2020; von Sawitzky et al., 2020).

Notification timing

Many on-bike sensor approaches issued notifications 4–5 s before a potential collision (Brummelen et al., 2016; Jin et al., 2021; Kalaiselvan, 2021). The system in (Engbers et al., 2018) provided warnings 3 s prior to a hazard. Assuming that C-ITS could theoretically enable earlier notification, they considered quite short timings of 3 s in advance (Oczko et al., 2020; Strohaeker et al., 2022). Such short notices leave cyclists with limited options, such as abrupt braking or abrupt avoidance manoeuvres, e.g. as observed in (Strohaeker et al., 2022) study based on cycling behaviour data. An earlier notification time of 9 s prior was used by (von Sawitzky et al., 2022), inspired by ISO 18682 referring to in-vehicle hazard notifications, which can be classified as warnings (1–3 s before an imminent hazard) or as awareness messages (AMs, 3–10 s before a potential hazard). About 64% of their participants found this early notification appropriate for dooring hazards, while 23% would prefer slightly later notification.

Notification content

Some approaches indicated a hazard without providing further information (Erdei et al., 2021; Jin et al., 2021; Manz et al., 2023; Prati et al., 2018; Strohaeker et al., 2022; Trösterer et al., 2022). Most approaches notified only about one hazard; four addressed multiple hazards (Engbers et al., 2018; Matviienko et al., 2018; Springer-Teumer et al., 2023; Strohaeker et al., 2022). Two systems specified which hazard to anticipate (Garmin Ltd, 2024; von Sawitzky et al., 2021). Both provided early notifications, which allow cyclist to notice and process the additional information. Several concepts included directional information (Brummelen et al., 2016; Engbers et al., 2018; Garmin Ltd, 2024; Matviienko et al., 2018; Oczko et al., 2020; Springer-Teumer et al., 2023; Vo et al., 2021). We assume that directional cues could facilitate the recognition and processing of potential hazards. Additional distance information was considered byvon Sawitzky et al. (2022). Some of their participants reported difficulties in mapping the indicated distance to accurately locate a parked vehicle. We assume that while distance information may be useful for static hazards, e.g. a parked vehicle, distance information may become ambiguous and add complexity when the hazard is in motion, e.g. a vehicle approaching from behind. Vo et al. (2021) used tactile cues to encode distances (far and close). Time information to indicate the approach of/to a potential hazard was suggested byvon Sawitzky et al. (2022), which in our view may be easier to understand than distance cues, as they indicate how quickly a hazard is approaching. Additionally, Matviienko et al. (2022) and von Sawitzky et al. (2020) provided further visual cues to aid in crossing decisions.

Feedback modality

Participants in the studies by Oczko et al. (2020), Trösterer et al. (2022), and (Sawitzky et al., 2022) did not prefer the use of uni-modal visual feedback alone over the alternative cues investigated. Oczko et al. (2020) found that visual-only cues led to significantly more collisions than auditory or haptic cues. Their participants rated the latter two cues as more effective in preventing collisions but perceived haptic feedback as easier to understand, less distracting, and this type of feedback also resulted in shorter reaction times. Trösterer et al. (2022) reported that visual-only cues on cycling glasses were difficult for their participants to perceive, e.g. in bright light conditions. Additional auditory cues could help to mitigate this problem. The analysis of cycling behaviour by (Sawitzky et al., 2022) suggested that uni-modal visual cues on cycling glasses led to similar evasive manoeuvres as with multi-modal visual and auditory cues. However, user experience ratings and follow-up interviews revealed a preference for a combination of visual and auditory cues. Some approaches that supported cyclists at intersections solely investigated visual augmented reality (AR) overlays (Matviienko et al., 2022; von Sawitzky et al., 2020). It was not reported whether participants disliked receiving only visual cues. However, participants of von Sawitzky et al. (2020) stated that an additional acoustic signal would improve the noticeability of the overlay. Visual cues alone could easily be perceived as distracting (Erdei et al., 2021; Matviienko et al., 2019), mainly because cyclists rely heavily on the visual (Mäkelä et al., 2015) channel. Also the auditory channel is required to detect or localize road users (Stelling-Kończak et al., 2016), e.g. to hear approaching vehicles. Therefore, some approaches evaluated only tactile cues (Brummelen et al., 2016; Engbers et al., 2018; Manz et al., 2023; Vo et al., 2021). For (Erdei et al., 2021)'s participants, auditory cues had a higher urgency than tactile or visual cues. Their participants also had difficulty perceiving/noticing visual cues, or perceived them with a delay. Auditory and tactile cues were perceived as equally comfortable and more comfortable than visual cues. The preferred uni-modal feedback was auditory (about 60%) or tactile (about 40%). However, 38% felt uncomfortable using auditory cues if others nearby could hear them. Strohaeker et al. (2022) suggested that combining the two modalities could mitigate missing cues due to ambient noise or ground vibration.

Few studies have explored multimodal hazard notification for cyclists. (Matviienko et al., 2019) examined combining (i) uni-modal cues (visual, auditory or tactile) indicating the direction of a potential hazard in combination with (ii) bi- or tri-modal feedback (combinations of auditory, tactile and visual cues) indicating imminent danger. They found multimodal signals conveyed higher urgency, which we see suitable to convey warnings. However, the combination of visual directional cues and a bi-modal warning indication which also included visual feedback were more distracting. Tri-modal cues were the preferred combination to warn about an imminent hazard. Springer-Teumer et al. (2023) also evaluated a tri-modal approach. Other studies focused on visual and auditory feedback combinations (Garmin Ltd, 2024; Prati et al., 2018; Trösterer et al., 2022; von Sawitzky et al., 2022). The participants in (von Sawitzky et al., 2022)preferred visual cues combined with auditory cues over visual cues alone. Trösterer et al. (2022) reported that their participants preferred auditory feedback alone rather than in combination with visual cues or visual cues alone. They also mentioned visibility problems in bright light conditions, which might have influenced their preference.

Notification presentation

The patterns and intensities of the signals varied between approaches. If patterns were described, we listed them in Table 4 in the appendix. Tactile cues were provided via handlebar grips (Brummelen et al., 2016; Erdei et al., 2021). Engbers et al. (2018) additionally used tactile feedback on the saddle. Vo et al. (2021) explored integrating tactile feedback into cycling helmets. Auditory cues were mostly provided by smartphones mounted on the handlebars (Strohaeker et al., 2022). The emitted sounds may be perceivable by others in the vicinity which may be uncomfortable for the user (Erdei et al., 2021). Alternatively, directional loudspeakers at ear level (Matviienko et al., 2019; von Sawitzky et al., 2022) or bone conduction headphones (Trösterer et al., 2022) have been suggested that offer more private and comfortable feedback. While mostly beeps were used (Erdei et al., 2021; Matviienko et al., 2019; Strohaeker et al., 2022; von Sawitzky et al., 2022), voice messages were also deemed helpful (von Sawitzky et al., 2022), albeit potentially distracting. Visual cues were presented through lights on the handlebar (Oczko et al., 2020), smartphones or other displays (Garmin Ltd, 2024), and smart glasses (Matviienko et al., 2022; Trösterer et al., 2022; von Sawitzky et al., 2020; von Sawitzky et al., 2022). Delivering visual information at head level, e.g. via head-mounted displays, may reduce distraction compared to handlebar-mounted displays, allowing users to obtain them at a glance or from their peripheral view (Matviienko et al., 2019). Smart glasses may also facilitate perception of the visual cue (Erdei et al., 2021; Strohaeker et al., 2022).

Design of the visual feedback

We used a screen-fixed 2D representation showing the information in the user’s direct line of sight (centre of the screen) to be visible independent of viewing direction. The direction of a hazard is indicated in a radar-like manner (circle segment), with a white icon depicting the kind of hazard in the centre. We anticipate that icons allow cyclists to quickly grasp which hazard to expect. A filling-up and colour-changing (green-yellow-dark orange) circle surrounding the radar shows the elapsed approximation time. We expected this filling circle to be easily visible in the cyclist's peripheral vision. If a hazard, such as an approaching vehicle, slows down during an active notification because it cannot overtake, the time display will remain at the lowest TTC previously recorded. This informs the cyclist that the hazard is still present, and that the alert is still active. When a hazard (is) passed, the visual fades out over a duration of 3 s. The visual was designed to be as transparent as possible by using lines thick enough to see the content but thin enough not to block the view. Visual cues were displayed at full brightness, but this was not sufficient outdoors. Thus, we placed a motorbike visor over the HoloLens 2’s projection area on the test track study. The visual representation was projected at a 2 m distance as recommended by Microsoft (2021) and took up about half of the horizontal screen size. Figure 2 shows the visual concept (a) and its temporal sequence (b).

Design of the supporting auditory or tactile cues

The supplementary feedback was provided at distinct times (9, 6, and 3 s before being close to a hazard and 1 s after passing it, inspired by the concept of (von Sawitzky et al., 2022) to convey the temporal approach.

For the auditory feedback, the cues varied in pattern and length. The same sinusoidal tones as in Sawitzky et al. (2022) that we describe in Table 1 were used.

The tactile cues additionally indicated the direction in which the hazard is located (relative to the own heading), similar to the audio-tactile guidance method in narrow field of view AR displays by Marquardt et al. (2020) (but they indicated absolute direction by head orientation). For the provision of tactile cues on the head, Vo et al. (2021) suggested using four tactons in a helmet to mediate cardinal directions, while (Krauß et al., 2021) proposed a headband with 12 tactons evenly distributed for directional cues. They used a vibration frequency of 200 Hz, which they reported to be slightly higher than recommended by Oliveira et al. (2016), to make sure the cue is perceivable. As our headband had to be worn with the HoloLens2 (see Figure 1 a), space was limited, so we placed six tactons to indicate front right, right, back right, back left, left, and front left directions. We housed the tactons in a foam headband and thus had to use a higher frequency as the 32 Hz recommended by Myles and Kalb (2010) for tactons placed directly on the scalp. We used the maximum recommended frequency of 150 Hz (Myles & Kalb, 2010) and vibration durations of 200 ms. The tactile cues had the same duration, except for the terminat cue. Their intensity increased (up to maximum frequency) when getting closer to a hazard. The end cue made a circular motion around the head. A Wemos D1 Mini Wi-Fi module (with ESP8266 microcontroller) was used to control the headband. Straps with side-release buckles allowed adjusting it to different head sizes.

Study 1 (Bicycle simulator)

This study was conducted in a Cave Automatic Virtual Environment (CAVE), a four-sided (left, right, front, bottom), 3×3×3 m-sized projection room, in which we placed a low-entry bike in a roller trainer facing the right front corner (see Figure 3 a). This allowed participants to shoulder-check approaching traffic. Bike movement was calculated based on the handlebar angle (360° angular sensor) and cycling speed (Bluetooth IMU measuring angular velocity in °/s).

The simulation environment was created in MathWorks RoadRunner (Mathworks, 2022). It was set in a suburban area, on a two-lane track (width: 3 m, each) with sidewalks (1.2 m) and a length of about1.4 km. Vehicles drove at 50 km/h and slowed down to the cyclist's speed when oncoming traffic delayed the overtaking manoeuvre. Traffic sounds were simulated so that participants could hear the approach of vehicles. Cyclists were overtaken with a lateral distance of 1.5 m (distance between the vehicle side and the end of the cyclist's handlebar). Vehicles spawned dynamically based on the participant’s cycling speed to ensure that participants encounter similar situations. One overtaking vehicle and up to three oncoming vehicles were spawned for each overtaking task. Participants encountered five overtaking vehicles per ride.

Study 2 (Outdoor test track)

We used the same bike on the outdoor facility (see Figure 3 b). It was equipped with a GeneSys SP80 GNSS receiver to obtain differential GPS position and cycling speed (at 20 Hz). The track consisted of two lanes (width: 3 m, each), a parking lane (2.4 m), cycling (1.5 m), and pedestrian path (1.2 m). On the track, 5 vehicles were parked in the parking lane. For each scenario, the ride consisted of 12 laps á 350 m (about 4.2 km in total). Cyclists rode next to the parking lane about 1.4 km.

On the computer used for the simulation of the scenario (triggering of the notifications), the software also displayed the connectivity states of the devices. In case of a disconnect, the experimenter could then stop the participant and manually reconnect the service, e.g. restarting the dGPS tracker. This setup also enabled the experimenter to monitor the notification states and which vehicle the participant would be notified about in real-time so that the experimenter could signal when doors should be opened. Another person switched between the parked vehicles in a fixed order (per Hazard Notification) when the participant returned to the start of the lap and was not able to see this person. When no vehicle was to be occupied, this person hid behind the test track barrier. Potentially dangerous situations were reduced by considering occupied vehicles only in two-thirds (8/12) of the laps and opening vehicle doors only for half of them (4/12) and only when the cyclist was still further away. The experimenter was on the phone with the person switching vehicles and gave a signal when a door was to be opened by counting down from three when the cyclist was 9 s away from the vehicle. Thus, a door opened about 5 s prior to reaching it. Doors were only opened when deemed safe by the person in the vehicle. The lap was repeated if doors remained closed due to the safety measure. The ride ended when the experimenter stopped the participant.

Communication between components

Scenarios and hazard notifications were developed in Unity (Unity, 2024). The primary simulation communicated with CAVE and HoloLens 2 (Google Remote Procedure Calls and ProtoBuf, gRPC (The Linux Foundation, 2024)), headband (TCP), angle sensor (USB) and IMU (Bluetooth).

Study 1 (Bicycle simulator)

Participants cycled along the main street and encountered oncoming and approaching vehicles. Each ride took about 6 minutes, and the overall study 60 to 90 minutes to complete.

Study 2 (Outdoor test track)

Participants rode alternating on the street or cycling lane next to parked vehicles (indicated by arrows on the HUD). Each ride took 20 to 25 minutes, the overall study 120 to 140 minutes.

Participants

In total, we could recruit 53 participants (details in Table 2) via university mailing lists and by word of mouth. For the bike simulator study, we asked people who are sensitive to motion (e.g. on boats, in cars, on amusement rides, or in VR) not to participate in the study. Due to the use of Microsoft HoloLens 2 glasses during both studies, participants with large-frame glasses could not participate unless they wore contact lenses. There were no other inclusion or exclusion criteria. Participants were not compensated for their participation.

Effect of Hazard notification

From a first visual inspection of the data presented in Figure 4, we could derive that user experience (pragmatic, hedonic, and overall quality) was rated rather high for both notification concepts. Workload ratings for rides with and without hazard notification appear similarly low. Further, effortlessness and intuitiveness of noticing a hazard were rated rather higher for rides with a system than unassisted rides (middle range). Ratings for gut feeling were in the middle range. Perceived safety for unassisted rides was relatively high but the rating was slightly higher for assisted rides.

For the factor Hazard Notification, there is evidence for equivalence, i.e. that it did not impact the ratings on hedonic quality, and workload, but the evidence was only anecdotal for overall quality and gut feeling. For the other measurements we found evidence for difference. Thus, we applied post-hoc Bayesian t-tests to gain further insights. For effortlessness ratings the tests provide extreme evidence that the ratings for no feedback (B) were lower than with V+A (PO = 508555.112), also when compared with V + T (PO = 169.528). There is anecdotal evidence that the ratings for V + T were equal to V + A (PO = 0.425). For intuitiveness, there is extreme evidence that V + A is rated higher than B (PO = 2217.207) and very strong evidence that V + T is rated higher than B (PO = 36.839). Ratings for V + A and V + T showed moderate evidence for equivalence (PO = 0.228). For perceived safety, the ratings for V + A (PO = 92.397, very strong evidence) and V + T (PO = 5.366, moderate evidence) were higher as for B. Also, V + A and V + T were rated equally with moderate evidence (PO = 0.199).

Effect of Hazard

For the factor Hazard, there was evidence for equivalence for all measurements. However, for intuitiveness and perceived safety this effect was only anecdotal.

Effect of Hazard notification × Hazard

Except for anecdotal evidence for difference of intuitiveness ratings, there was evidence for equivalence for the other measurements and the interaction effect did not impact the ratings.

Effect of AMs

Additional support was perceived as always helpful for VRUs (5). About 67% of participants found AMs valuable and helpful in enhancing awareness of potentially dangerous situations (36). AMs would improve their situational awareness by being more alert (15) and increase their sense of safety (8). Notifications further facilitated detecting approaching vehicles, requiring less concentration (5), and allowing time for preparing to act (12). Some individuals expressed reluctance to use AMs (6). They did not want to be disturbed (1), distracted (1), did not see an added value (1), or stated to be capable of navigating traffic safely on their own (3).

Perception of AMs

Timing. Most participants found receiving notification 9 s before encountering a hazard suitable (31, about 85%); however, some perceived the notification as too early but could imagine being notified 5–6 s ahead of a potential collision (12, about 22%).

Sensory perception of cues. After the familiarization ride, we asked participants, whether the provided cues were perceptible and clear. All responded in the affirmative, apart from one participant in the simulator study with very curly hair, who perceived the tactile feedback but would have appreciated a stronger feedback intensity.

Visual cues. Icons were perceived as an effective way to indicate potential hazards (37). For some participants, the visual cue was distracting, confusing due to too many factors displayed, or hard to focus on (5). For others, the direction indicator clashed with the time indicator as they moved in the same direction (3). For some (3), the visual blocked their view.

Auditory cues. Compared to tactile cues, they were more comfortable (9), increased attentiveness (4), were more intuitive (4), more unambiguous (2) and easier to notice and understand (3). Further, they were familiar from games (1) or navigation (2). However, some participants felt detached from the situation (3) or distracted (4) or found it annoying or uncomfortable (2). Particularly in noisy environments (5) or when listening to music (2) sounds could be hard to notice. The initial sound was startling (5), or disturbing (3). Multiple occurrences of auditory cues could reduce attention paid to the AMs (1).

Tactile cues. Compared to auditory cues, tactile cues were perceived as more comfortable (9), like a nudge (1), not distracting (2), better comprehensible (4), less intrusive (1), more subtle (3) and made more aware (3). Tactile cues were easier to notice (also in noisy environments) (2) and allowed for easier hazard localization (3). However, they were also perceived as confusing (5), ambiguous (4), strange (2), distracting (4), irritating (9), startling (1), or too intimate (1). Some found the end notification too strong (11), or vibrations felt uncomfortable (2). A prolonged use could induce headaches (1).

Proposed adjustments

Visual cues. Some participants liked the presentation of the visual cues and saw no need for further improvement (23). For some only sounds or vibrations would suffice to be more alert (5). It was suggested to place the visual off-centre (6) or on the screen borders pointing in the direction of the hazard (2). Position, scale, and transparency should be configurable (5). Further, content visibility could be increased with thicker lines (5). The icon should indicate vehicle types (3). To better distinguish the direction and time indications, a horizontal progress bar could be used (2). The time indicator could be replaced if the color-coding is transferred to the direction indicator (3). Red-green colour blindness should be considered (1). Additionally, it's transparency could be reduced after a few second to be less present in the field of view (1).

Supporting cues frequency. Auditory cue volume should be adjustable (5). The end vibration could be replaced with a sound (1) or two short vibrations (2). Some stated they would not need an end notification (13). Others would only require an initial indication to be made aware of the notification (6). It was suggested to provide the vibration away from the head on the handlebar (1) or through a bracelet (1). Vibrations should be distinct, as the tones were (1), and their intensity should be adjustable for varying hair density and improved perception (5).

Design ideas for presenting repeating occurrences

Visual indicators. Participants suggested indicating multiple occurrences near the hazard icon, either signalling the number of vehicles (14) or that there will be more encounters (5). The icon could display multiple vehicles (6), or multiple instances of the visual could be stacked (1). Alternatively, the direction indicator could mimic browser tabs (4) or stacked from the inner to the outer area (4). They could also point toward the direction the hazard will be the closest to the cyclist (4) or their initial direction (2). The time indicator could either disappear after the occurrence of the first hazard (3), stay filled (5), update for subsequent vehicles (4), or overlay a second indicator (9). Also, time information could be coded into the direction indicators (1).

Tactile and auditory indicators. Our participants could imagine continuing the same cues as for the intermediate steps and a terminating cue (9) or only using start and end notifications (2).

What to notify about

AMs should inform about potential dooring situations (19), about approaching cars, at intersections during crossing and left or right turns (19), about vehicles taking off (parking space or driveway) (5) and when cyclists are in a blind spot (2), or nearing accident black spots (1). It would be helpful to notify about e-vehicles (7), e-scooters or e-bikes (3) and other cyclists in their vicinity, e.g. when they are behind them and could overtake (4), about ghost riders (3) or when there is a high number of cyclists (2), crossing pedestrians/children (14). Some want to be notified about ground conditions, e.g. road surface, grooves, potholes (6). It was also proposed to include navigation (3), weather forecast (2) and traffic lights (5) notifications.

AM support could be unnecessary when cycling on fields/in forests (5), when encountering obvious dangers (4), in calm areas (2), in very crowded areas (2) or on distinct and broad cycling lanes (3). It would likely be impossible to notify about unpredictable situations, such as small objects on the roads or the movement of children, pedestrians, or animals (3).

Challenges

Participants highlighted the necessity of a longer adaptation period to grasp the notification system fully (17). Some participants noticed a learning effect over time (6). However, some participants reported concerns about over-reliance on AMs, as they quickly relied on the AMs and paid less attention to their environment (15), which could become dangerous (1). One participant suggested shifting the visual cue to the side to avoid solely focusing on it.

Additional remarks

AMs would benefit those wanting to be safe in traffic (1) or children and older people (5). They would also be beneficial to cyclists listening to music by either using a V + T concept (2) or integrating V + A and overlaying the music with cues (2). V + T could be particularly useful for hard-of-hearing cyclists (1). The notification device should be lightweight and comfortable (8), preferably integrated into helmets (3) or glasses (8). Due to vibrations (ground, cyclist movement), the displayed content bobbed slightly up and down; also, the glasses may slip from their initial position, reducing the content visibility (1). The system should be configurable (4), and also notify about immediate hazards (3). When encountering hazards simultaneously, the notifications should not be overwhelming (6); prioritizing hazards could help in such a case (1).

Design challenges

The intended purpose of AMs is to inform cyclists in advance of potential hazards. The underlying algorithm can only detect foreseeable potential hazards. It is unlikely that it will predict spontaneous hazards. For example, dooring can happen when a vehicle stops briefly at the roadside, and someone exits (Johnson et al., 2013). Cyclists might still face hazards that the system cannot foresee, even if it can provide notification about this specific kind of hazard.

Further, in each user study, we focused only on one hazard and only one incident at a time. However, it is likely that multiple hazards occur in close sequence or also simultaneously. One participant proposed that it may be required to prioritise some hazards over others. Being overtaken and dooring situations sometimes overlap. In such cases, either one type of hazard must be prioritised, or both hazards must be indicated. Both approaches come with their own set of advantages and disadvantages. When prioritising, the most urgent/relevant hazard could be specified, but other hazards are disregarded. The indication of multiple simultaneous hazards at once provides a more comprehensive overview of potential hazards but could also lead to distraction and overload as a lot of information needs to be displayed.

Challenges for the use(rs) of AMs

One major concern stated by our participants was that users may become overly reliant on these notifications, an issue also raised by Matviienko et al. (2022); Springer-Teumer et al. (2023) and (von Sawitzky et al., 2022), though not reported by their participants. This problem of overreliance is common to support systems, e.g. advanced driver assistant systems (Geels-Blair et al., 2013; Wintersberger et al., 2019). Further research should focus on preventing hazard blindness to (other) hazards due to overreliance and ensuring that users to stay alert. Further, some of our participants expressed a desire for a longer familiarisation phase with the notification. The potential of such a phase to help cyclists understand and learn the correct use of AMs and their limitations needs to be further explored in future work.

Design recommendations for early hazard notifications for cyclists

We derived the following recommendations for the design of hazard notification that are provided on head-mounted systems, such as smart cycling glasses. The following list is not a comprehensive set of recommendations, but it reflects those derived from the research presented in this paper.

Recommandation 1. Sufficient notification time should be allocated to allow users to notice and understand the notification, to identify the hazard, to consider which countermeasure is appropriate, and act if necessary. 

While many of our participants found a timing of 9 s ahead to be adequate and stated that this time frame allowed them to prepare to react to a hazard, some would have preferred a later indication at 5-6 s prior. A timing of about 6 s prior before a potential collision (Gold et al., 2013) may still allow sufficient time for situation processing and reaction. Although further research is needed to examine the impact of shorter notification times, we assume that too short time frames could lead to overwhelm. Users may not be able to process the provided information and also react to the potential threat. Other studies on cyclist hazard notifications (Engbers et al., 2018; Oczko et al., 2020; Strohaeker et al., 2022) considered shorter time frames but provided less detailed information about the situation. For a system designed to support multiple kinds of hazards arising from interactions with motorists, we anticipate that providing more detailed information is crucial to enable fast and intuitive interpretation. Therefore, we assume that ensuring sufficient time for processing and acting is essential.

Recomandation 2. Unambiguous icons or symbols should convey information about the kind of hazard the notification addresses. 

Many participants found the icons indicating the type of hazard to be an effective way to communicate potential dangers. We assume that using distinct icons can minimize the misinterpretation of hazards, further users likely become familiar with their meanings quickly.

Recommandation 3. Bi-modal cues should be used. While relevant information should be provided visually, another modality should be used to guide the cyclist without requiring them to focus on the visual content. Here, the peripheral field of view should also be utilized.

Both bi-modal concepts that our participants experienced were perceived as useful. Few participants indicated that guiding cues alone would be sufficient for notification. However, we suspect that it may be difficult to identify hazards, when visual information about the hazard is not indicated.

Recommandation 4. The used head-mounted device should be comfortable to wear.

Comfort is crucial for user acceptance and long-term use of the notification device. Cyclists should not be distracted or irritated by wearing the device during their rides.

Recommandation 5. The notifications should be customizable in a predefined extent.

Participants expressed a desire to customise AM presentation according to their preference, such as adjusting volume, intensity, or when notifications are issued. Allowing some customization ensures that cyclists can adjust the notifications to their needs. This customization also includes that the type of additional feedback modality (auditory or tactile) can be selected by the user, as some of our participants stated to rather not use the support if it is not their preferred feedback type.

Parallels to existing HMI guidelines. Naujoks et al. (2019) proposed a checklist of 20 design guidelines for human-machine interfaces (HMIs) in automated vehicles, taking into account existing guidelines, design recommendation and best practices in this context. While some of these guidelines are not relevant to notifications for cyclists—being more focused on the functionalities of automated vehicles, such as operation principles or system mode indication—two closely resemble our design recommendations for cyclist hazard notifications. For example, guideline #9 suggests using ‘[c]ommonly accepted or standardized symbols [(such as pictograms)] […] to communicate the automation mode’. Similarly, our recommendation #2 suggests using icons to convey the hazard type. Additionally, guideline #6 advises that ‘[t]ime-critical interactions with the system should not require continuous attention’. Although this guideline differs slightly from our recommendation #3 that guiding cues should be used to allow cyclist to maintain their visual focus on the road, both emphasize minimizing the need for frequent interaction with the system.

We argue that some of the other guidelines could also be considered as recommendations for cyclist hazard notifications, e.g. guideline #16 suggests ‘[a]uditory output should raise the attention of the driver without startling her/him or causing pain’. Relating this guideline to our proposed notification, this advice would apply to both supporting cues (auditory, and tactile) that we investigated. While not all of (Naujoks et al., 2019)'s guidelines are directly applicable to cyclist hazard notifications, they offer valuable reference points for future research aimed at developing guidelines for cyclist notification on smart glasses.