A mobile measurement system from DLR was used for data collection (see left columns of Figure 1). It allows to continuously record traffic data during day and night-time under various weather conditions. Two sensor systems are mounted on a telescopic mast, which can be extended to a length of up to 8 meters. Each sensor system consists of a stereo camera for vision-based traffic data collection and an infrared flash. The latter enables to operate in adverse lighting conditions, including dusk and night. The server infrastructure inside the trailer processes the images captured by the sensor system and detects, tracks and classifies motorized road users (MRUs) and VRUs, producing anonymized trajectory data and video material. Motorized road users are classified as either car, van, truck or other narrow vehicles including motorbikes. VRUs are classified as either cyclists or pedestrians. However, due to technical challenges during the measurement campaign, pedestrians could not be detected robustly. Consequently, the following analysis about VRU and MRU interactions includes cyclists only, with narrow vehicles excluded to minimize ambiguity. The trajectory data contains information on the position in Universal Transverse Mercator (UTM) coordinates, velocity in km/h, orientation in degrees, and dimensions (length, width, height) of road users in meters. The video material is monochrome and has low resolution, which enables traffic situation reconstruction, while license plates and faces remain unrecognizable.

The two selected measurement locations are situated on the Federal Road 179 in Brandenburg, Germany, south of Berlin (see Figure 1). We will describe both measurement sites in more detail in the following.

To analyze the behavior of traffic participants at both measurement locations, the traffic data provided by the measurement station is semi-automatically processed according to specific scenarios. In the following section, we will explain the methods used for scenario identification, extraction and analysis, also referred to as scenario mining by (Gelder et al., 2020). The methods used are chosen based on the specific needs of each measurement site, which we present separately.

During the campaign at the measurement site Bestensee, a total number of 74,231 motorized road users (MRUs) and 516 vulnerable road users (VRUs) were recorded. Similar numbers were recorded for the measurement site Pätz with 78,087 MRUs and 281 VRUs. An overview of the overall daily traffic distribution for both campaigns is shown in Figure 5. The number of MRUs for the first measurement campaign Pätz is shown in the first row and for the second measurement site Bestensee in the second row. The traffic distribution is aggregated by the time of the day, while night spans six hours and all other times of the day three hours with morning starting at 6 a.m.

The daily overall traffic volume at the measurement site Pätz remains at a similar level, but with an increased volume in the (late) afternoon on Sundays. The first also applies to the measurement site Bestensee, but with an overall decreased volume on Sundays. The higher traffic volume on Sundays at the measurement site Pätz may be attributed to tourist traffic to the lakes, such as the Pätzer Vordersee. Video material shows vehicles parking on the road in the west of the intersection that leads to the lake (see Figure 6).

Figure 5 also shows that both measurement sites have a similar traffic volume distribution per day, except for Sundays. This is despite the fact that the first campaign took place at the beginning of July and the second at the end of September. This similar traffic volume may be attributed to commuter traffic, as both measurement sites are only about two kilometers away from each other. Worth noting is the difference in traffic volume on Tuesday of the second measurement site Bestensee, which is not the case for Pätz. This may be attributed to a public holiday on 03 October 2023, and thus a lower commuter traffic.

An overview of the mean traffic volume per hour is given in Table 1 for the different classes of traffic participants on weekdays. The majority of traffic participants are cars for both measurement sites, followed by vans for the site Pätz and trucks for the site Bestensee. It is worth noting that the highest traffic volume per hour for cyclists is observed at Sunday for both measurement sites. For the site Bestensee, the difference between Sunday and Monday, and Tuesday is only margin. This may be attributes to the holiday in the second week.

For both measurement sites, we also estimated the associations of traffic participants to routes. The traffic volume per route is shown in Figure 7. The two routes with the highest traffic volumes are SouthNorth and NorthSouth on both measurement sites. At the measurement site Bestensee, there is only a low traffic volume on the two other secondary routes. This is different at the measurement site Pätz with nearly a third of the overall traffic volume on the two routes NorthWest, WestNorth.

A more detailed overview of the daily traffic distribution on the two primary routes over the course of the day for both weeks is illustrated in Figure 8. The profiles represent the mean number of traffic participants per hour and day, with the shaded areas illustrating the 95% confidence interval. The first row shows the number of traffic participants for the measurement site Pätz and the second for Bestensee. The NorthSouth route has its peak on weekdays in the afternoon for both measurement sites and at mid-day on the weekend. The SouthNorth route's peak on weekdays is in the late morning, whereas the peak is higher for the measurement site Bestensee than for Pätz.

The difference in the course of the number of traffic participants between weekdays and weekends may be attributed to commuter traffic. Traffic participant may use the SouthNorth route to commute to work before noon and the route NorthSouth to commute home after noon. To assess whether the route, daytime and site have an effect on the traffic volume, a negative binomial generalized linear model was fitted. We chose the model because traffic volume data often exhibits overdispersion. The Negative Binomial distribution has been demonstrated to be a robust method for analyzing such data, providing a more accurate fit than a standard Poisson model. The model was estimated based on 1,536 observations for weekdays, with each observation being the average traffic volume per hour. It should be noted that observations from Tuesday were excluded from analysis due to the public holiday that occurred during the measurement period at Bestensee. Given the focus on commuter traffic, the observations are grouped by daytime. The morning period is from 05:00 to 10:00, and the afternoon period from 14:00 until 20:00. The model demonstrates a strong fit, with a Pseudo R² = 0.712, indicating that approximately 71.2% of the variance in traffic volume is explained by the predictors.

Table 1 presents the results of four distinct hypothesis t-tests based on the fitted model. Each test examines the difference in traffic volume between morning and afternoon for specific site and route combinations. The coefficients represent the estimated difference on a logarithmic scale between morning and afternoon traffic, with negative values indicating higher afternoon traffic volume. The findings indicate that for the NorthSouth direction, traffic volume is significantly higher in the afternoon than in the morning at both Bestensee (β = −0.8677, p < 0.001) and Pätz (β = −0.8373, p < 0.001). However, this correlation is not observed in the SouthNorth direction. There is no significant difference in traffic volume between morning and afternoon. Thus, it can be concluded that the hypothesis of commuter traffic is only partially confirmed. There is a significant increase in traffic volume on the NorthSouth route during the afternoon compared to the morning. But, there is no evidence to suggest that SouthNorth route experiences commuter traffic during the morning.

Due to the rural characteristics of the measurement site Bestensee, we analyzed the velocity of traffic participants travelling out of town and into town. The velocity distribution of the MRUs on the main routes at the measurement site Bestensee is illustrated in Figure 9 indicating that the velocity depends on the traffic participant type and the travelling direction. Overall, traffic leaving the town is slightly faster than traffic entering the town with a mean velocity of 54.32 ± 9.47 km/h and 53.25 ± 9.32 km/h. The majority of traffic participants (64.85%) to not adhere to the speed limit of 50 km/h, with 21.74 % driving faster than 60 km/h and 4.91 % faster than 70 km/h.

To assess whether there is a significant effect of the traffic participant class and direction on velocity, a two-way analysis of variance (ANOVA) was conducted using ordinary least squares (OLS) models. To evaluate the significance of each effect, Type II sum of squares were applied to account for interactions between direction and class. The fitted OLS model explained 2.2% of the variance in velocity (R² = 0.0366, adjusted R² = 0.036). Type II ANOVA revealed significant main effects of direction (F(1, 65068) = 218.32, p < 0.0001) and class (F(2,65068) = 532.70, p < 0.0001), as well as a significant interaction between direction and class (F(2, 65068) = 572.60, p < 0.0001). The findings indicate systematic differences in mean velocity across groups, with the interaction suggesting that the effect of direction varies by traffic participant class. Despite these highly significant effects, the relatively low R² indicates that most variance remains unexplained. This may be due additional unmeasured factors.

Separate hypothesis tests were conducted to examine whether the change in velocity between the SouthNorth and NorthSouth directions was statistically significant within each vehicle class (see Table 2). The coefficients represent the estimated velocity difference between SouthNorth vs. NorthSouth, with negative values indicating lower velocity at the SouthNorth route.

The findings indicate that cars travel significantly but only marginally faster (Cohen’s d = 0.19) on the NorthSouth direction than on the SouthNorth route (β = 1.78, SE = 0.076, p < 0.001). The opposite holds for trucks (β = −8.03, SE = 0.288, p < 0.001) and vans (β = −3.25, SE = 0.556, p < 0.001) travelling significantly slower on the SouthNorth direction than on the NorthSouth direction, while the effect is stronger for trucks (Cohen’s d = 0.86) than for vans (Cohen’s d = 0.35). It can thus be concluded that the effect of travel direction on velocity is moderated by vehicle class, with cars driving marginally faster on NorthSouth direction, whereas vans and trucks drive significantly slower.

The results thus far showed that the traffic at the measurement site Bestensee was mainly observed on the federal road and only a minority on the turning roads. In this work, traffic participants are assumed to be in a crossing scenario, if their trajectories geometrically intersect and their absolute PET is lower than ten seconds. At the measurement site Bestensee, we found only a few scenarios meeting this criterion. However, t the measurement site Pätz we found a vast number of crossing scenarios. Due to this, we will focus on the crossing scenarios observed at the measurement site Pätz.

Traffic participants following the WestNorth route may interfere with those that are on the NorthSouth route because they need to cross their path to turn. The left panel of Figure 10 shows the trajectories of traffic participants that turn left to merge onto the federal road from west. The figure shows a significant variation in the turning behavior of traffic participants. Those with a lower curve radius tend to steer towards the lane marking of the federal road, while those with a higher curve radius follow a more direct path to merge onto the federal road.

The right panel of Figure 10 also represents this variation in turning behavior by showing the crossing points of traffic participants involved in a crossing scenario. Crossing incidents are represented by blue color for higher numbers and black for lower numbers. The crossing points are widely distributed along the federal road, which may be due to turning behavior. However, the figure also shows that the most of crossing points are concentrated in a smaller area located in the middle of the intersection, rather than directly in front of the road coming from the west.

To analyze the crossing scenario in terms of traffic safety, we estimate the PET according to (2) between the traffic participants that are involved in such a scenario. The top panel of Figure 11 shows the distribution of the signed PET value of all crossing scenarios (N = 3,886). The sign of the PET value indicates the order in which traffic participants leave the conflict area. A negative PET indicates that the traffic participant following the WestNorth route leaves first, while a positive PET indicates that it leaves second. The figure shows that more crossing scenarios were observed, where the traffic participant following the WestNorth route turns before the traffic participant on the federal road (NorthSouth route).

The result of classifying a crossing scenario according to the PET is shown in the bottom panel of Figure 11. The figure illustrates the cumulative distribution function of the PET for the two variants of crossing scenarios (before and after) separately. The PET of traffic participants turning first (before) is lower than turning second (after). Specifically, 3.76% of before and 0.93% of after have a PET lower than two seconds and 30.57% of before and 12.28% of after lower than three seconds.

We manually analyzed a portion of the crossing scenarios with a PET lower than three seconds using video material. Our analysis revealed that in several situations, an accident was avoided because the road users on the federal road waived their right of way and actively prevented a critical situation by braking. Figure 12 shows two examples of such crossing scenarios. The active rear braking lights indicate the braking in both situations.

At the Bestensee measurement site, we analyzed the behavior of traffic participants in overtaking scenarios. Our study focused on the distances between motorized road users (MRUs) overtaking each other and MRUs overtaking cyclists, as all traffic participants share the available infrastructure. During the measurement periods, we observed 399 overtaking scenarios. Out of these, 200 involved a MRUs overtaking cyclists heading into the village (on the NorthSouth route), while 129 involved a MRUs overtaking cyclists heading out of the village (on the SouthNorth route). Additionally, there were 15 scenarios of MRUs overtaking other MRUs on the NorthSouth route and 55 on the SouthNorth route. The mean lateral distance between the traffic participants during the overtaking scenario is illustrated in Figure 13. Kernel density estimation with a gaussian kernel was used to estimate the cumulative distribution functions (CDFs).

To assess the effect of direction and traffic participant class on the overtaking distance, a generalized linear model (GLM) with a Gamma distribution and log link function was fitted. The Gamma distribution was selected because of the shape of the CDFs illustrated in Figure 13. To assess whether the overtaking distances illustrated in Figure 13 are significantly different across direction and class, hypothesis tests were conducted (see Table 3). The analysis reveals that MRUs travelling out of town (SouthNorth route) maintained a significantly shorter overtaking distance when overtaking MRUs (β = −0.2131, SE = 0.062, p < 0.001) compared to travelling into town (NorthSouth route). Conversely, VRUs were overtaken with a significant but marginal increase in overtaking distance when travelling out of town compared to travelling into town (β = 0.0913, SE = 0.024, p < 0.001). Furthermore, the analysis indicates that MRUs were overtaken significantly closer than cyclists when travelling out of town (β = 0.2107, SE = 0.034, p < 0.001). In the opposite direction, the overtaking distance to cyclists compared to MRUs is marginally lower, though this effect is not statistically significant (β = −0.0937, SE = 0.057, p = 0.1).

Another perspective on the lateral distance in overtaking scenarios is given by Table 1. It shows mean lateral distance and standard deviation of the different overtaking scenario variants. The mean overtaking distance of cyclists on NorthSouth is µ = 2.36 m with a standard deviation of 0.51 m and on SouthNorth is µ = 2.59 m with a standard deviation of 0.52 m. The Table 1 also illustrates the probability P(D < x) of a lateral distance D being lower than a specific threshold x in meters for both routes, as well as for overtaking another motorized vehicle or a cyclist. It is worth noting that, of all scenarios where a cyclist was overtaken, 27% had a lateral distance of less than two meters on the NorthSouth route and 14.74% on the SouthNorth route, which is the minimum overtaking distance in extra-urban areas in Germany.

Figure 15 illustrates four examples of overtaking scenarios that were observed during the measurement period. The bottom right image shows an ambulance overtaking a car which makes way by driving on the grass shoulder. The other three examples show cars or a bus overtaking cyclist. It is worth noting that in the top right image, the car significantly undercuts the minimum safety distance early in the morning, despite no traffic coming from the opposite direction.

In the following, we will highlight various phenomena that we found during both measurement campaigns which potentially affect traffic safety.