In 2020, the inhabitants of Zurich, the most populous city in Switzerland, voted for a more bike-friendly infrastructure. The goal is for cyclists to be able to bike with priority on a network of 130 km throughout the town before the year 2030. However, one requirement was that most of the projects have to be implemented on the existing roads without increasing the road width or building new roads (Stadt Zürich, 2023).

Since 2020, various measures have been taken to improve the quality and safety of cycling, e.g. removal or relocation of parking spaces, signalizing routes, and making them visible with markings or changes to the right-of-way regulations at road junctions. On many sections of the preferential cycle routes, the speed limit is 30 km/h, and car traffic is reduced to increase cyclists' safety. In challenging sections, research is carried out on how this specific section can be improved to increase the safety of cyclists and their perceived safety. One of these challenging sections is addressed in the current study.

Cyclists must be and are deemed to feel safe to increase the share of cycling in overall traffic. However, since objective and subjective safety do not necessarily correlate (Elvik & Bjørnskau, 2005; Hackenfort, 2012), it is relevant to investigate both aspects before deciding on a specific design. Various studies indicate that the perception of safety is an essential factor regarding the social transition to a higher share of cycling in traffic, especially in emerging cycling cities (Heesch et al., 2011; Kaplan & Prato, 2016; Rondinella et al., 2012). If cyclists generally feel unsafe, they might bike less than they would under different circumstances (Götschi et al., 2016) or even develop cycling avoidance (Chataway et al., 2014). Overestimations of risk, i.e. people feel unsafe, although objectively it is safe, can cause people to be more or even too alert and attentive to the traffic around them. While these cases usually are not a safety problem, the reverse scenario is far more critical: a traffic spot is considered safe although being objectively unsafe; cyclists might then not pay as much attention as needed (Ghielmetti et al., 2017; Hackenfort, 2012; Klebelsberg, 1982; Petermann et al., 2008).

Cycling lanes are generally considered a good choice for reducing the accident rates of cyclists (Park et al., 2015; Pulugurtha & Thakur, 2015), especially if these lanes are wide enough for cyclists to overtake each other. Internationally, colored cycling lanes are used to draw attention of motorized traffic in potential conflict zones (Autelitano & Giuliani, 2021; Fyhri et al., 2021). Furthermore, in some countries, such as Switzerland or Belgium, the red color is used to remind motorized traffic that these areas are primarily for cyclists. However, Autelitano and Giuliani (2021) summarize that it takes more than coloring cycling lanes for cyclists to feel safe.

Accident statistics are often considered to determine the objective safety for cyclists on different infrastructures (Chen et al., 2012). However, at least in dedicated traffic spots, accidents are statistically rare and, therefore, not considered reliable (Smith, 1976). Additionally, accident statistics typically underestimate the accurate number of cycling accidents by far (Hertach et al., 2022; Winters & Branion-Calles, 2017) because less severe accidents or incidents are often not recorded by the police. Therefore, other traffic safety parameters are increasingly considered, such as near misses or conflicts (Steiner et al., 2023). Traffic conflicts are events that are sufficiently close to real crashes (Tarko, 2018). Video-based traffic analysis has become a standard tool for recording traffic conflicts (Beitel et al., 2018; Kronprasert et al., 2021). Nevertheless, the data is difficult to interpret due to a lack of standards and dependencies on road design (Steiner et al., 2023).

While there is ample evidence regarding the impact of adequate cycling infrastructure on increasing cyclists’ safety, the effects of specific cycling lane design on safety perception are understudied, as different researchers point out (Chaurand & Delhomme, 2013; Rietveld & Daniel, 2004; Schepers et al., 2014; von Stülpnagel & Binnig, 2022). Moreover, as infrastructure differs significantly between countries or even cities, so do influencing factors such as speed limits for motorized traffic (Pulugurtha & Thakur, 2015), traffic density (Parkin et al., 2007) or the behavior of motor vehicles (Henson et al., 1997). While the participants of a US study rate cycling lanes almost as uncomfortable as biking on a street without any cycling infrastructure (Foster et al., 2015), in a study carried out in Germany, participants rated cycling on cycling lanes far less dangerous than cycling on a street without a cycling lane (von Stülpnagel & Binnig, 2022).

von Stülpnagel and Binnig (2022) investigated which factors and which design combinations contribute to the subjective safety of cycling lanes. In an image-based online study, they manipulated different factors such as lane surface (green vs. same tarmac as the street), lane width, left cycling lane buffer, such as dashed lines, continuous lines, and physical separations. The authors found that structural separation is perceived as safer than cycle lanes without separation, which is in line with findings from Foster et al. (2015) and McNeil et al. (2015). Both studies collected their data in the US, which has a substantially different urban infrastracture than Switzeland has. In the latter, due to limited available space, the separation of traffic lanes for cyclists, tramways, motorized traffic, and pedestrians is not always possible. Therefore, a study based on a more condensed road design was deemed to be necessary.

Furthermore, like Foster et al. (2015), von Stülpnagel and Binnig (2022) collected the data based on pictures, which might differ from assessments that are collected in real traffic. For this reason, this study additionally aimed to compare the prospective assessment based on images with those assessments of the same situations in real traffic.

Therefore, in this field study, different designs of a cycling lane at a challenging road section are investigated regarding (1) their objective safety based on recorded conflict using video-based analysis, (2) the perceived, and (3) expected safety (based on pictures presented at the first interview).

Based on that, the following research questions were addressed:

1) Does light segregation with discs or guide beacons decrease the number of adaptive actions between cyclists and motorized traffic

2a) Are there any differences in the subjective safety assessments based on these conversions?

2b) Which form of separation is considered the safest by cyclists?

3) To what extent are the expected safety assessments based on pictures comparable to the perceived safety assessment when the conversions are in place?

In order to avoid conflicts of interest, the investigating scientists were not involved in the decision-making process of determining the area of interest. Three conversions were in place for one each month over the spring and summer of 2022. The three conversions were: (1) continuous line that is forbidden to cross for motorized traffic under penalty. Therefore, the continuous line was interrupted in two short sections to allow access to the parking garage. (2) Light segregation using additional discs. These elements can be driven over both by cyclists and motorized traffic. And (3) light segregation with guide beacons, which might break if motorized traffic hits them. The aforementioned conversions are shwon in Figure 2.

With respect to research questions 1 and 2, a between-subject design was applied with four experimental conditions. As dependent variables, the recorded conflicts were used to describe objective safety, while the interview questions were used to assess perceived or expected safety. Regarding research question 3, a within-subject design was chosen.

The interviews were conducted three to four weeks after the conversion so bicyclists could get familiar with the new design before they were interviewed. The video recordings took place outside of the interview times. Days with good weather conditions without rain were generally chosen to reach as many participants of the opportunity sample as possible.

The video recordings were carried out during all stages on a total of six weekdays (Monday to Saturday) with a focus on peak traffic times so that the evening peak of shopping traffic from Monday to Friday between 4 pm and 7 pm and the shopping traffic on Saturday between 9 am and 7 pm were covered. Two programmable recording devices collected the traffic movements in the entrance/exit area of the parking garage from different perspectives.

The video recordings of 9 hours per stage were evaluated quantitatively and qualitatively following each survey. The quantitative analyses included the determination of traffic volumes using semi-automated techniques. The qualitative analysis focused on the cyclist categorized into (1) the bike passed through without encountering a motor vehicle intending to cross the cycle lane. (2) The bike passed a motor vehicle that intended to cross or actually crossed the cycle lane while no adaptive action of the cyclist (e.g. braking, swerving) was necessary. (3) The bike passed a motor vehicle that intended to cross or actually crossed the cycle lane while an adaptive action of the cyclist (e.g. braking, swerving) was observed. The latter condition was interpreted as a conflict (see examples in Figure 3).

The city police supported the interviewers by stopping the cyclists. The cyclists' participation in the interviews was voluntary; the police officers were not involved in the questioning in any form. There were no exclusion criteria for participation, so every cyclist passing the site was stopped. All interviewers were wearing safety vests, which identified them as researchers of the university ZHAW.

Due to practical considerations, the questionnaire was kept as short as possible. Besides demographics, the participants were asked about their driving behavior, general safety perception as a cyclist, experience with the specific road section, and current safety perception. Additionally, during the first interview period, the cyclists were asked about the expected safety of the three conversions using a photo montage to answer research question three (see translated questionnaire in Appendix A). All safety-related questions were asked on a six-point Likert scale (1 ‘very unsafe’ to 6 ‘very safe’).

Differences in the objective safety in the form of comparing the recorded adaptive actions such as braking or swerving were analyzed with an ANOVA.

For the expected safety based on images, a repeated measurement ANOVA was used because all participants at the first interview rated all three conversions based on images.

An ANCOVA was used to investigate differences in perceived safety differences in real traffic. It is expected that general safety perception in Zurich and general safety perception on bicycle lanes influence the rating, which is why we added them as covariates.

A logistic regression was carried out to investigate factors influencing perceived safety, including the design. Therefore, the safety assessment as the outcome variable was converted into a binary variable, comparing those who rated the current situation as very safe (6 on the Likert scale) compared to the rest. This re-coding was chosen because very high scores were generally reached, and we wanted to distinguish those who felt very confident from the others. Moreover, this came close to a median split.

Video-based traffic analysis was applied to address the first research question of whether light segregation with discs or guide beacons decreases the number of adaptive actions between cyclists and motorized traffic. For the initial situation with the original colored bike lane (Figure 2, picture A), it was revealed that 20% of the interactions with motorized traffic were linked to an evasive action (Figure 4, Table 3). Exposure-relative, most of the cyclists' adaptive actions were seen when the motorized traffic turned right to enter the parking garage.

Results of an ANOVA show that the conversions, in general, led to significantly fewer recorded adaptive actions (F (3, 1 529) = 8.869; p ≤ .001). Due to a significant Levene`s test, the Brown-Forsythe correction was applied. Post-hoc tests were significant for all three conversions compared to the initial situation: continuous line (p ≤ .001), discs (p ≤ .001), and guide beacons (p ≤ .001). However, the three conversion conditions did not differ from each other.