How do cyclists experience a context-aware prototype warning system? Assessing perceived safety, perception and riding behaviour changes through a field study

How do cyclists experience a context-aware prototype warning system? Assessing perceived safety, perception and riding behaviour changes through a field study Georgios Kapousizis a,* , Annemarie Jutte b , M. Baran Ulak a , Karst Geurs a a Department of Civil Engineering, University of Twente, Enschede, the Netherlands b Ambient Intelligence Group, Saxion University of Applied Sciences, Enschede, the Netherlands ARTICLE INFO Keywords: E-bike Perceived safety Perceived trust Cycling behaviour Human-machine interaction Acceptance ABSTRACT The number of bicycle crashes is increasing in many European countries. In the Netherlands, a country wellknown for its high-quality cycling infrastructure and cycling culture, bicycle crashes are also increasing, especially for e-bike users. Smart bicycle technologies, such as safety warning support systems, could contribute to reducing crash risk for cyclists. However, perceived safety and trust in these technologies are determinant factors in accepting and using such technologies. This study investigated users’ perceived safety, perceived trust, and perceived performance with a context-aware prototype warning system to support cyclists in the real world. In addition, it investigated users’ riding behaviour changes when receiving warnings in high crash risk locations by collecting GPS data. The above were evaluated through a field trial experiment using three rides per participant, with the first one serving as a control ride and a follow-up questionnaire after each ride conducted in Enschede, the Netherlands, between April and May 2024, with a sample of 46 participants. Results show that participants’ perceived safety increased after they tried out the prototype warning system. In addition, it was found that warning systems positively influence participants’ riding behaviour, since they reduced their speed. This study proves the potential benefits of smart bicycle technologies in improving cyclists’ safety. 1. Introduction The number of bicycle crashes and fatalities increases yearly worldwide (European Commission, 2023a; ITF, International Trasnport Forum, 2023). In Europe, every year, around 2000 cycling fatalities happen, with no decrease since 2011, representing 10 % of all road fatalities (European Commission, 2023a, 2023b). This is probably due to the increased number of people cycling in the last few years and exposure to motor vehicles in many countries with and without cycling culture (European Commission, 2023b; Uijtdewilligen et al., 2022). Moreover, when we account for road fatalities in urban areas, cyclists and other vulnerable road users, such as pedestrians, represent almost 70 % of the total fatalities (European Commission, 2023b). The majority of road fatalities within urban areas involve motor vehicles with cyclists and pedestrians. The recent report from the International Traffic Safety Data and Analysis Group (IRTAD) shows a discouraging trend in cycling fatalities, especially for e-bike users. The share of e-bike users in cyclist fatalities has been increasing in many countries, such as Switzerland (55 %), Germany (44 %), Belgium (38 %), and the Netherlands (34 %) (ITF, International Trasnport Forum, 2023). In the Netherlands, well known for its cycling policies, culture, and infrastructure, cycling fatalities represented 40 % of all road fatalities in 2022, and the e-bike users rate of all cycling fatalities increased by 9 % in 2022 since 2018. This is mainly due to the e-bike users’ exposure, age and health factors (Westerhuis et al., 2024). In addition, since 2020, the number of bicycle fatalities in the Netherlands has been higher than that of passenger cars (Statistics Netherlands (CBS), 2023). Therefore, additional actions should be taken to decrease cycling crash risk and bring these numbers down. 1.1. Background Smart bicycle technologies have rapidly developed over the last decades, aiming to influence cyclists’ safety and the general view of bicycles, e-bikes and the future of soft transport modes, especially in the * Correspondence to: Department of Civil Engineering, University of Twente, Drienerlolaan 5, Enschede 7522 NB, the Netherlands. E-mail address: [email protected] (G. Kapousizis). Contents lists available at ScienceDirect Journal of Cycling and Micromobility Research journal homepage: www.sciencedirect.com/journal/journal-of-cycling-and-micromobility-research https://doi.org/10.1016/j.jcmr.2024.100051 Received 12 September 2024; Received in revised form 26 November 2024; Accepted 28 November 2024 Journal of Cycling and Micromobility Research 3 (2025) 100051 Available online 2 December 2024 2950-1059/© 2024 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). era of Internet of Things (IoT) and smart cities (Behrendt, 2016; Oliveira et al., 2021). These technologies and systems have various functionalities that potentially impact road safety (Engbers, 2019; Westerhuis et al., 2021). Smart bicycle technologies may attract more people to cycling since they could offer a sense of increased perceived safety, as well as an actual increase in cyclists’ safety. Smart bicycle technologies can assist cyclists in reducing the risk of crashes, collisions with other vehicles, and single-bicycle crashes due to the wide range of applications. However, users’ perception of smart bicycle technologies and trust will influence their deployment and penetration rate into the market. Recently, considerable literature has grown on warning systems aiming to prevent collisions (Kapousizis et al., 2022). Linder et al. (2024) examined users’ perceived safety and the effectiveness of a bicycle warning system in reducing crashes and human interaction with autonomous vehicles in different scenarios and intersections using a bicycle simulator. Engbers (2019) examined how technologies, such as frontand rear-view assistance on bicycles, can support older adults by conducting experiments on bicycle simulators and on the road. The findings of this study showed that older adults valued these systems positively, especially the rear-view assistant. In addition, Westerhuis et al. (2021) tested a light communication system to support older adults while cycling. The aim was for the system to display braking and turning intentions, hence, users were aware of the direction of the bicycle and able to press a button to indicate their direction rather than using their arms, which influences cyclists’ balance, especially for older adults. Participants of the experiment, both users of the instrumented bicycle and users riding other non-instrumented bicycles who interacted with the system, found the system useful. Another study that investigated users’ perspectives of a warning for critical events using a bicycle simulator was conducted by Kreißig et al. (2022). They developed hypothetical critical safety events and warned the participants in a virtual environment. The assessment of this study revealed good ratings regarding trust and acceptance of the warning system and higher acceptance for the young participants (Kreißig et al., 2022). Other studies investigated an intersection safety application alert system in field trials using an on-site camera capturing traffic and identifying conflicts based on time to collision (TTC) and post encroachment time (PET) measurements (Soro et al., 2024). When both TTC and PET values were less than 3 seconds, a warning was sent out to the cyclist approaching the intersection. The results of this study showed that participants were willing to use the safety application and preferred to receive simple and clear warnings (Soro et al., 2024). Similarly, Schories et al. (2024) examined the performance of a warning triggers system for cyclists based on a TTC warning with 4 seconds. They applied this method in a virtual simulation, and the results showed positive safety benefits of avoiding a collision with a vehicle. Lastly, Hagelen et al. (2019) and Husges and Degen (2021) used radar and lidar sensors on the bicycles to scan the surroundings and estimate TTC and, in case of an imminent collision, send warnings to the cyclists. The findings of these studies indicate the potential benefits of such systems and sensors on cyclists’ safety. To increase cycling safety, researchers also designed and investigated the effectiveness of external nudging approaches aiming to adjust cyclists’ behaviour (Fyhri et al., 2021; He et al., 2019; Kovaceva et al., 2022; Wallgren et al., 2020). While a viable option, this type of feedback is more rigid than having an easily adjustable on-bicycle system. However, previous research on-vehicle safety communication has mostly focused on drivers (Biondi et al., 2017; Grushko et al., 2021; Spence and Ho, 2008). Drawing from the literature on drivers, Biondi et al. (2017) found that multimodal audio-tactile warnings decreased the braking reaction time more than single modal warnings in the simulation environment. However, Geitner et al. (2019) found that audio-tactile multimodal warnings were perceived as being more startling. Strohaeker et al. (2022) studied the effectiveness of different audio and tactile modalities in different conflict scenarios. They found that audio warning was more effective since participants had shorter reaction time compared to the tactile. Thus, in situations where the users need an urgent reaction, it is preferable. Tactile communication was found to serve better in less urgent situations (Strohaeker et al., 2022). Erdei et al. (2020) found that cyclists most consistently responded to audio signals across different real-world environments compared to visual and tactile signals. Tactile signals performed significantly worse on rough roads. However, Erdei et al. (2020) did not consider how users perceived the different signals, which will be explored in this research. Recently, the municipality of Amsterdam conducted a field experiment in order to test users’ reactions when they received warnings (Tollenaar and Plazier, 2023). They used multimodal audio-visual communication, with the visuals serving to inform the cyclist of speed warnings, and the audio serving to make participants aware of the system. However, participants mentioned low satisfaction with the manner of communication (Tollenaar and Plazier, 2023). For the effectiveness of any warning system, perceived safety and trust are major factors (Jahanshahi et al., 2020; Nordhoff et al., 2020; Simsekoglu and Kl¨ ockner, 2019); therefore, it is important to test users’ perceptions and preferences towards new technologies. Moreover, Kapousizis et al. (2024) found that perceived safety has a major role in user behavioural intention to adopt a technology, especially in the Netherlands. Kapousizis et al. (2024) also indicated that testing a smart e-bike in a real environment would capture users’ preferences better, although several studies used online surveys to investigate user’s preferences for an on-bicycle warning system. For instance, using an online survey, De Angelis et al. (2019) examined user preferences for warning systems with audio-visual or haptic modes; however, they did not find a significant difference between them. Nonetheless, Kapousizis et al. (under review-a, under review-b) found that such warning systems have potential and users are willing to pay for them. This means that bicycle users are open to use warning systems to improve their safety. Although there is evidence in the literature for the perceived safety and preferences of cyclists towards warning systems, the majority of existing studies used hypothetical scenarios and surveys. Soro et al. (2024) recently utilised field trials to examine perceived safety in intersections and imminent collision scenarios. 1.2. Research objective The studies above provide valuable insights into different types of warning systems and the importance of user acceptance, however, more research is needed to examine users’ opinions and acceptance of a specific context-aware cyclist warning system. This study aims to develop and test a context-aware prototype warning system to support e-bike users in high crash risk locations. For this purpose, we designed a field experiment, developed a warning system, identified high crash risk locations, designed a survey, and devised a data collection plan. The objectives of this study are as follows: 1. To investigate cyclists’ perceived safety, trust and overall perception of the context-aware cyclist warning system. 2. To investigate the effect of users’ riding behaviour receiving warnings from the warning system. 3. To evaluate cyclists’ preferences for different communication methods and test whether the audio or haptic may offer a more intuitive interaction with the users. The developed warning system falls to the bicycle smartness level 2, “warning assistance”, based on the classification proposed by Kapousizis et al. (2024) and Kapousizis et al. (2022), which also has the highest technology readiness level. In addition, Kapousizis et al. (2024) mentioned the importance of users trying out different systems since it might influence their opinions and attitudes. Therefore, alternative communication methods were tried to identify the approaches which cyclists find valuable and motivating while not being too distracting or startling. G. Kapousizis et al. Journal of Cycling and Micromobility Research 3 (2025) 100051 2 The rest of the paper is organised as follows: Section 2 describes the methodology, field trials, system development, and data collection approach; Section 3 presents the collected data and analyses of the results; Section 4 discusses the results, and Section 5 presents the conclusion of this paper. 2. Methodology 2.1. Fieldwork set-up The methodology followed in this paper is divided into six parts: 1) the experiment design of the field trials, 2) the safety system design, 3) the cycling crash density model, 4) the apparatus, 5) participants recruitment, and 6) the questionnaire and collected data. Fig. 1 illustrates these parts and serves as the conceptual model of this study. In detail, we designed a field experiment and used bicycle crash data to model cycling crash density. We used bicycle crash data and the road network to model cycling crash density to identify high crash risk locations using the Kernel Density Estimation (KDE) (please refer to Section 2.2). Then, we used the results of the crash density model to feed into a smartphone application to design a system to warn users via visual, audio, and tactile (vibration) notifications, which consists the context-aware warning system. In addition, we developed a survey distributed throughout the field trials to evaluate the context-aware safety support system and users’ perceived safety, trust, perception, and riding behaviour changes. The results from this study could be used to further improve the system, as indicated in Fig. 1. Overall, we organised field trials in Enschede, the Netherlands, including various types of infrastructure (mixed traffic, dedicated bicycle paths/lanes) to investigate how users react to receiving notifications and examine their perceived safety. 2.1.1. Field trial experiment design A smartphone-based application was developed and was fed with high crash risk locations as identified based on the crash density model described in the following section. The aim was for participants to ride an instrumented e-bike on a predefined route and receive warnings when they approached a high-risk location to pay more attention and reduce their speed. We selected a route around the University of Twente for the field trial. The reason for this was twofold: 1) practical, we were able to host participants in the university’s room, and also the bicycles were there, and 2) we were able to identify a route which was 3.4 km long with different types of infrastructure (mixed traffic, dedicated bicycle paths and bicycle lanes), with a shopping area, school zone; in a total of five high crash risk areas (four locations, one was ridden two times). Fig. 2 shows the route used for the field trials. Participants were requested to ride the same route three times. During the first round, which served as a baseline ride, the set-up operated normally, but no information regarding road safety was given. This means the warning system was effectively turned off; only the speedometer was shown on the smartphone, as in the ‘Approaching’ situation in Fig. 3. The function of this round was to 1) get participants familiar with the new e-bike and 2) to understand the perception and perceived safety of participants when riding this specific route with a “conventional e-bike”. During the second and third ride, the warning system was turned on. Participants received warnings through either the audio or tactile modalities combined with visual information each time. In the third round, participants received warnings through the modality they had not experienced in round 2. In the high crash risk locations, participants were given an advisory speed of 20 km/h; this choice was based on pilot studies during which 18 km/h was found to be much too slow, matching the findings of Tollenaar and Plazier (2023), where participants indicated that advisory speeds of 10 or 15 km/h were generally too low. Navigation instructions were given to the participants through the Komoot (https://www.komoot.com/) on one of the smartphones. In detail, participants rode the e-bike prototype in a predefined route. During the first ride, participants rode the prototype without receiving warnings, that is namely, as a “conventional e-bike”. However, additional sensors such as GPS were used to collect speed data without influencing their riding behaviour. After the first ride, participants filled out the first part of the survey, where they stated their perceived safety, perceived trust and experience. We repeated the survey about users’ perceived safety, trust, and experience when they received notifications to reduce their speed. Participants rode the prototype for the second and third ride, receiving warnings from the context-aware prototype warning system based on different modalities. Through this, we investigated how users felt when receiving notifications from the warning system, whether they were annoyed or comfortable, and their perceived safety for the context-aware warning system. Overall, we examined to what extent users were blissful as well as their riding behaviour. Furthermore, data was collected using GPS from the smartphone, and other sensors were incorporated into the e-bikes, adding valuable characteristics in understanding participants’ riding behaviour. The GPS, which was collected during the rides, allowed us to examine to what extent users followed the notifications and advice the contextaware prototype warning system provided. 2.1.2. Crash density Bicycle crashes have been studied widely in the literature, and there are different approaches to doing so. We aim to identify high crash risk locations in Enschede and develop a safety support system using bicycle crash data. Many studies have used Kernel Density Estimation (KDE) to identify traffic crash clusters (Abdulhafedh, 2017; Anderson, 2009; Ulak Fig. 1. Conceptual model of the study. G. Kapousizis et al. Journal of Cycling and Micromobility Research 3 (2025) 100051 3 Fig. 2. Crash density map of Enschede and location of field trial route. Fig. 3. Interaction flow of the context-aware warning system prototype. G. Kapousizis et al. Journal of Cycling and Micromobility Research 3 (2025) 100051 4 et al., 2017). KDE is a non-parametric approach to estimating the intensity of a spatial process, which focuses on clustering pattern distributions throughout an area study region and creates a density surface of spatial point events over a 2-dimensional geographic space (Xie and Yan, 2008). However, in traffic crashes, the events occur in a network; thus, the KDE may not be a good fit (Xie and Yan, 2008). To this end, we used the Network Kernel Density Estimation (NKDE) since road crashes occur along a network, and the density needs to be estimated in a 1-D linear space rather than the Euclidean distance commonly used by the KDE. The NKDE is an extension of the classical KDE. We used the package spNetwork developed by Gelb (2021) in R Core Team (2023) to proceed with the estimation, which is free and open-source software. We used the road network from the OSM (OpenStreetMap, 2023) and police crash report (BRON) data from 2016 to 2019. Fig. 2 illustrates the bicycle crash patterns in the municipality of Enschede resulting from the NKDE; lines with red represent road segments with high crash density. In addition, we used OSM data in order to identify school zones and high street locations, which were found to have a significant impact on bicycle crashes (Kapousizis et al., 2021). 2.1.3. Safety system design The context-aware warning system proposed in this paper communicates the potentially dangerous locations to cyclists. The prototype of the warning system is created as a smartphone app using Android Studio. The context-aware warning system consists of two levels of communication: notifying and informing, similar to the set-up by Tollenaar and Plazier (2023). The context-aware warning system comes into action when a ‘transition event’ occurs, and an overview of the event loop can be found in Fig. 3. Four such events are identified: 1) entering: the cyclist enters a dangerous zone; 2) exiting: the cyclist exits a dangerous zone; 3) speeding up: the cyclist speeds up from below the advisory velocity to above; and 4) slowing down: the cyclist retains a ‘safe’ velocity. Additionally, the entering event is split up into two cases: the cyclist is going above or below the advisory velocity. Notifications are either given as audio signals through boneconducting headphones or tactile signals through haptic gloves. Boneconducting headphones have been chosen since they do not block any environmental sounds. Haptic gloves were chosen due to their ease of implementation, as opposed to haptic handlebars (Baldanzini et al., 2011). To minimise distraction, notifications are only sent when potential danger to the cyclist increases during the entering and speeding up transition events. Additionally, the system aims to supply less intrusive information that provides additional understanding for the notifications. The aim is to present this information in a way that is easily accessible and understandable without providing too much distraction. The information is given through a smartphone screen as shown in Fig. 3, similar to the implementation of Tollenaar and Plazier (2023). The smartphone screen is set to red when the cyclist is going above the advisory speed and is set to green when the cyclist is going below the advisory speed. Only the speedometer is shown on normal segments without increased safety issues. 2.1.4. Apparatus Participants were riding a smart e-bike prototype on a predefined route starting at the University of Twente, the Netherlands. The field trial - set up/data collection/recruitmentwas a joint work of PhD students as part of the Smart Connected Bikes project (https://www.smar tconnectedbikes.nl/). Therefore, the prototype bike incorporated multiple sensors to collect data for different research projects. Not all collected sensor data is used in this paper. For this research, the e-bike was equipped with two Nokia C32 TA1524 smartphones. One of the smartphones served as a prototype of the context-aware warning system and for GPS measurements; the other was used for navigation. Participants were asked to wear bHaptics TactGlove DK2 haptic gloves or Shokz OpenMove S661 bone conduction headphones for the warning system during the second or third round. Additionally, for related research, one forward-facing GoPro 3 action camera and three Inertia ProMove Miniwere IMU sensors were mounted on the bicycle. Two of the three IMU sensors were attached to the bicycle frame to measure road surface quality, one to the pedal. In addition, participants were asked to wear a helmet to which another Inertia ProMove Miniwere IMU sensor was attached. Participants also wore an Empatica E4 wristband and a Polar H10 heart rate sensor. Finally, participants were also asked to use a button press system to indicate their experience during the rides. The audio signal consists of two 780hz signals of 300 ms with a 20 ms pause. This is the same frequency is as used by Erdei et al. (2020). They used two signals of 650 ms instead of 300 ms. However, the shorter duration is chosen since it can be argued that an (unnecessarily) long signal can be more distracting. The phone volume was set to 50 %. Following Erdei et al. (2020), the tactile signal consists of two pulses of 250 ms with a 400 ms pause, however, it is not possible to extract the frequency from the bHaptics interface. The bicycle was equipped with different hardware features, such as two smartphones, one to give directions to users and the other to communicate safety warning messages, as indicated in Fig. 4. 2.1.5. Participants recruitment The field trial, along with a questionnaire, was conducted between April and May 2024 in Enschede, the Netherlands. The questionnaire was developed using Maptionnaire (Burnett et al., 2023) and was filled out by the participants during the field trials. The target population were people 1) older than 18 years, 2) who have been cycling on average at least once per month in the past six months, 3) without the influence of illicit drugs, and 4) without medication that is issued with advice against, or prohibition against, participating in traffic. We used different channels to recruit participants, such as mailing lists of the University of Twente and Saxion University of Applied Sciences, advertisements in online press (UToday), local cycling unions in the city of Enschede, portal news of the municipality of Enschede, the Enschede Fiets app and personal social media. The field trials with the questionnaire had a completion time of around 1.5 hours. Each participant received a voucher worth € 10 as a token of appreciation for their time. In total, 46 participants were recruited and participated in the field trials. Before the field trials started, initial pilots with the instrumented ebikes took place to test that all sensors worked properly and that the setup was feasible. In addition, we tested the survey; we first piloted it Fig. 4. Bicycle set-up. G. Kapousizis et al. Journal of Cycling and Micromobility Research 3 (2025) 100051 5 within the research group and later with some participants. Once the pilot of the field trials and the survey were successful, the survey was translated into Dutch by native speakers within the research group and tested once more to ensure the readability and consistency of the translation. 2.1.6. Questionnaire Enschede has a population of around 150 thousand inhabitants, with a dense cycling infrastructure as do most Dutch cities. The city of Enschede has international influences since the University of Twente is located there, and there are many international and high-tech companies. Thus, the survey was translated into Dutch and English to account for both native and foreign citizens. Note that the survey was developed in English, and two pilots were used to ensure the optimal structure and reliability of the questions. Thus, the questionnaire was distributed among researchers from our groups, and after two iterations, we translated the survey into Dutch. In total, 46 participants joined the field trials. After cleaning the data, 41 were analysed; two participants did not complete the rides due to rain, and three participants experienced technical issues, so we decided to exclude them. All participants completed three rounds: 25 completed the second round using audio warnings and the third round with the tactile, while 16 completed the second round with the tactile communication and the third round with audio warnings. The number of experiments that resulted from audio and tactile warnings was merely due to the availability of the equipment. Questions in the survey for this research were divided into two parts: 1) questions related to sociodemographic statistics, participants’ general mobility habits, and participants’ perceived safety and experiences when they cycle in their daily lives; 2) questions about the participants’ opinions on the context-aware warning system. Note that the second part was repeated three times, once after each ride. Participants received a written information brief about the steps of the field trials as well as an oral explanation. Initially, participants were requested to fill out the first part of the survey, which consisted of sociodemographic information as mentioned above. Then, participants were requested to proceed by riding the bicycle. During this ride, no warnings were given through the prototype. Once the ride was completed, participants filled out the second part of the survey. Then, participants received an explanation of the safety support system and a set of 5-Likert scale questions about their expectations of the system. The explanation of the context-aware warning system given to the participants during the survey was as follows: “In our research, we want to evaluate the design and effects of a warning system. The system warns you to avoid collisions with other road users. The warnings are given at hazardous locations, through visual, audio and/or vibration signals. We would like to ask you some questions about your expectations of this system.” Then, participants continued with the second and third rides, during which warnings were given when entering critical locations. Table 1 presents an overview of the questions we used in the survey. 2.2. Research methods In this study, we had two main sources of collected data: 1) survey data and 2) GPS data, which was collected using smartphones. Participants’ answers to the survey give us information about their perception of the context-aware warning system. At the same time, GPS data allows us to assess the effects of the system based on different modalities on participants’ riding behaviour, specifically in speed reduction. We calculated the average speed of the participants 5-seconds before and after the point they were supposed to receive a warning. The GPS used during the field trials had 1 Hz accuracy, recording 1 point per second. After testing the context-aware support system, we used a 5-second window as the proper time to allow participants to react after receiving a warning. Thus, we used 5-second windows before and after high crash risk locations. Initially, we used t-tests to examine whether speed reductions in high crash risk locations differed by personal characteristics, such as gender, and usage of e-bike and conventional bicycle and if they were statistically significant. We used 299 speed points for the three rides for the “SlowDown” event to examine whether participants reduced their speed when initially receiving the warnings. We further examined users’ speed reduction after receiving system warnings by applying a Multinomial Logit model (MNL) to the collected data. We refer readers to the following study for more information about the MNL (McFadden, 1973). We estimated two MNLs to analyse participants’ speed when they approached a high crash risk location and their speed when they received warnings to reduce their speed. For this part of the analysis, we used the “entering”, “slow down”, and “slowdown reminder” since these are the cases in which participants receive notifications. In total, we used 1560 warnings, 524 warning points for audio communication, and 592 for tactile communication. The remaining 444 data points are from the first ride (baseline) participants completed and were supposed to receive warnings. In order to perform the analysis, we used the five-second average speed before and after the warnings. 3. Results 3.1. Data statistics As mentioned earlier, 41 participants were included in the analysis after the data cleaning. The mean age of participants was 42.1 (SD = 15.65) and ranged from 25 to 75 years of age; 73 % were males (n =30), and 63 % (n =26) earned a university degree. In addition, 73 % (n =30) of the participants use a bicycle weekly, more than four days a week, while only 22 % (n =9) use an e-bike weekly, meaning they have a quite good cycling experience. None of the participants used a SpeedPedelec in a related question we asked. The sample distribution can be Table 1 Overview of the questions used. Category Question Source Perceived safety I would feel safe with the Warning System. (Jahanshahi et al., 2020; Kapousizis et al., 2024) and own investigationI think with the Warning System I can increase my safety I think that riding with the Warning System can reduce the risk of me getting involved in a crash compared to a conventional bike I think that there will be fewer crashes for users with the Warning System. Perceived trust I would trust the Warning System (Hinderks et al., 2018; Kapousizis et al., 2024; Nordhoff et al., 2020; Venkatesh et al., 2012) I will ride with more stress using the Warning System. I would like to use the Warning System. I think the Warning System would be easy to use. I expect that the capabilities of the Warning System will meet my requirements. Perceived performance Useful to useless (Hinderks et al., 2018) and own investigationAssisting to worthless Undesirable to desirable Understandable to badly understandable Raising alertness to reducing alertness Noticeable to badly noticeable Motivating to Very demotivating G. Kapousizis et al. Journal of Cycling and Micromobility Research 3 (2025) 100051 6 found in Table 2. 3.2. Perceived safety and trust Participants’ scores for perceived safety and trust based on a 5-Likert scale of questions are shown in Figs. 5 and 6. Participants completed a repeated survey after every ride (without intervention, with audio and visual, and tactile and visual). Thus, we were able to capture participants’ opinions about the context-aware warning system. As shown in Fig. 5, question Q1: I would feel safe with the Warning System indicated that participants felt safer with the warning system compared to the baseline, especially with tactile communication. The second question, Q2: I think the Warning System I can increase my safety, shows that participants had higher expectations since the score for the baseline has the highest value, while the audio and tactile come slightly after. In addition, we notice that while there was 22 % “Neutral” in the baseline, later, after the rides, participants shifted to the negative side and felt less safe with the tactile warning. Regarding Q3: I think that riding with the Warning System can reduce the risk of me getting involved in a crash compared to a conventional bike; we see that for the baseline, 44 % of participants believed that the Warning System would decrease their risk of involvement in a crash, and 34 % was neutral. However, after the rides, the neutral percentage shifted to the negative side since the percentages were 27 % for the audio and 34 % for the tactile. This means that participants were not fully convinced that the warning system could increase their safety compared to a conventional bicycle. For the last question, Q4: I think that there will be fewer crashes for users with the Warning System; 49 % of the participants had a neutral opinion, 44 % positive and only 7 % negative for the baseline. However, we see that after using the warning system, participants have a positive opinion in this regard, especially for the tactile warning. Overall, we see that during the first round of questions (before the participants experience the system) it seems that they keep a neutral attitude, while this changes after the rides. In addition, we see a shift from neutral to negative, implying that participants were less satisfied. However, we also found that participants had a positive opinion about the context-aware warning system for three out of the four questions. Regarding participants’ trust for the context-aware warning system, as shown by Fig. 6, participants’ scores for Q1: I would trust/trusted the Warning System, are in the middle, with 37 % indicating a positive opinion for the baseline. After participants rode the e-bike with the warning system, their trust reduced to 23 % for the audio warnings; for the tactile, their trust remained at 37 %. However, most participants had a negative opinion in both audio and tactile settings, indicating a low trust in the system. Regarding Q2: I will ride/rode with more stress using the Warning System; for the baseline, the majority of the participants expected to feel lower stress. However, they mentioned that they felt higher stress using the Warning System, especially for the tactile. About question 3, Q3: I would (like to) use the Warning System, most of the participants have a negative opinion, with only 21 % and 28 % having a positive opinion of the audio and tactile, correspondingly. Q4: I think the Warning System would be/was easy to use, we see that the majority of the participants have a positive opinion, with 79 % thinking that tactile is easy to use and 67 % for the audio. It is important to mention that before the rides, 44 % of the participants had a neutral opinion, which changed after the rides. Regarding the last question, Q5: I (expect) that the capabilities of the Warning System will meet/met my requirements, the positive answers have remained similar between the baseline and the rides; however, half of the participants initially had a neutral opinion (before they try out the e-bike), they shifted to negative ones. 3.3. Perceived performance During the survey and field trials, participants were also asked to evaluate the warnings they received on a number of scales. The resulting scores the participants gave can be found in Fig. 7 and Fig. 8. Participants were asked for this evaluation after rides 2 and 3. As can be found in Fig. 7, most participants indicated that they found both the audio and tactile warnings noticeable. As many as 97 % of the participants found the tactile warnings noticeable, while 73 % found the audio warning noticeable. The scores for alertness are mostly towards raising alertness, although some participants did feel like the system reduced their alertness (51 % for audio and 53 % for tactile). Regarding the scores for motivating the system, most participants had a neutral attitude, while only 32 % had a positive opinion for the audio and 36 % for the tactile. Lastly, participants evaluated to what extent the warnings were understandable, with most participants (78 %) indicating that they found the tactile clear, while the audio was only 49 %. Participants’ opinions on the context-aware warning system’s usefulness and desirability varied (Fig. 8). In detail, 36 % of the participants found the tactile warning useful, and 41 % found the audio useful. Regarding the system’s desirableness, 29 % of the participants gave a positive score for the audio and tactile aspects. However, we noticed that a high proportion of the participants had a neutral opinion for both questions. 3.4. Riding behaviour changes First, we compared participants’ speed for the baseline ride (without intervention) against the rides with the intervention. We applied the same for the rides with the interventions. We calculated the average speed for all the locations, before and after warnings, and found that participants reduced their speed by 1.3 kilometres per hour after intervention. In addition, we ran t-tests in order to examine the difference in Table 2 Sample composition. Variable Sample Count Percentage Number of respondents 41 100 % Gender   Male 30 73 % Female 11 27 % Age   <25 4 10 % 25–35 18 44 % 36–45 2 5 % 46–55 7 17 % 56–65 5 12 % 66–75 4 10 % >75 1 2 % Education   Low (high school or lower) 6 15 % Vocational (Technical) 8 20 % High (university degree or higher) 26 63 % Other 1 2 % Net monthly individual income ( € /month)  up to 3000 26 63 % More than 3000 15 37 % E-bike usage   Never or less than 1 day per year 25 61 % 1–5 days per year 6 15 % 6–11 days per year 1 2 % 1–3 days per month - 0 % 1–3 days per week 2 5 % 4 days or more per week 7 17 % Conventional bicycle usage   Never or less than 1 day per year 6 15 % 1–5 days per year - 0 % 6–11 days per year 2 5 % 1–3 days per month 3 7 % 1–3 days per week 4 10 % 4 days or more per week 26 63 % G. Kapousizis et al. Journal of Cycling and Micromobility Research 3 (2025) 100051 7 average speed between the two groups. In detail, we used the ride without intervention as a baseline and compared it with the audio and tactile interventions (Table 3). The average speed for all three rides was almost the same before a warning was communicated to the users (baseline: 22.9 km/h, audio: 23.1 km/h and tactile: 23.0 km/h). After the warning, users reduced their speed on average by 1.2 km/h and 1.4 km/h for audio and tactile, correspondingly. Fig. 9 shows the average speeds before and after the warnings for the individual modalities, audio and tactile, in the average 5-second windows before and after participants enter a high crash risk location, and the average speeds for both modalities. Furthermore, as mentioned earlier (Section 2.1), in our experiment, we used a route with five different high crash risk locations (one was a school location), where participants were receiving warnings. Thus, we also compared participants’ speeds before and after warnings in the different locations. We compared the participants’ speeds when entering a location without and with the context-aware warning system, and we found that participants had lower speeds after receiving warnings when they entered high crash risk locations, and the speed reduction was almost equal in all locations. 3.5. Factors affecting speed reduction Table 4 presents the test results based on different types of warnings and the different variables that were used. Males were found to have a higher speed reduction than females in both types of warnings. For the audio, males had a reduction of 1.2 km/h (p-value <0.001) and females only 0.6 km/h; similar rates were found for the tactile 1.3 km/h (p-value <0.001) for males and 0.9 km/h for females after they received the warnings. The differences in speed reduction were not significant for female participants. Females were found to cycle with an average speed of 21.8 km/h before receiving a warning, while males 23.5 km/h; females used to cycle 1.7 km/h slower than males. Therefore, both males and females reduced their speed; however, speed reduction for the males was higher since they generally approached a critical location with a higher speed. The last variables we tested were the usage of bicycles weekly or not and types, such as e-bikes and conventional bicycles. We tested participants using an e-bike on a weekly base (more than one time per week) against those who do not use an e-bike weekly. We found that both groups reduced their speed by 1.1 km/h after receiving audio and 1.4 km/h when receiving tactile warnings. However, only the nonFig. 5. 5-point Likert scale questions related to perceived safety. G. Kapousizis et al. Journal of Cycling and Micromobility Research 3 (2025) 100051 8 weekly e-bike users were found to have a significant difference. This is probably due to the small number of weekly e-bike users. For the conventional bicycle users using a bicycle weekly, it was found that a speed reduction of 0.9 km/h for audio and 1.3 km/h for tactile warnings were both significant, with a p-value <0.05. However, for the other group, participants using a bicycle less than once a week, were found to have non-significant differences again due to the small number of participants in this group. We also considered controlling for multiple corrections by using, for instance, the Bonferroni correction; however, since this study was exploratory and restricted to a few, not simultaneous, comparisons, we did not use any corrections (Armstrong, 2014; Lee and Lee, 2018; Streiner and Norman, 2011). Furthermore, we estimated a MNL and used the following categories: •0: participants did not follow the recommendation; •1: participants reduced their speed up to 1 km/h; •2: participants reduced their speed between 1 and 2 km/h; •3: participants reduced their speed by more than 2 km/h. We estimated two models, one for the audio warnings and one for the tactile. Table 5 and Table 6 present the results correspondingly. During the model estimation, we used the category “0: no speed reduction” as a Fig. 6. 5-point Likert scale questions related to trust. G. Kapousizis et al. Journal of Cycling and Micromobility Research 3 (2025) 100051 9