1 Introduction

While current advanced driver assistance systems (ADAS) have already improved the safety and comfort of manual driving [5], automated driving is expected to lead to even more benefits, such as reduced congestions and an increased mobility for a large number of people [18, 33]. However, there are a range of human factors issues that need to be overcome prior to launching automated vehicles [27].

When used in urban traffic, automated vehicles (AVs) need to be able to interact with vulnerable road users (VRUs), such as pedestrians and cyclists [13]. Since interaction in road traffic as a social phenomenon can be complex and can involve ambiguities [21], a detailed understanding of how human drivers and VRUs interact in road traffic is essential for the development of appropriate algorithms for AVs [16, 20]. In addition, cooperation as a collaborative effort is required for a successful interaction: In a joint action, road users share common goals and thus follow common solutions instead of interacting competitively [9, 11, 15]. Therefore, it is crucial to examine how goals and intentions of VRUs can be recognized and how to communicate AV’s intentions to them. It is assumed that such a cooperative interaction between AVs and VRUs can lead to higher satisfaction, trust, acceptance, efficiency and safety in road traffic [12, 15].

While the previous project KIVI investigated the cooperative interaction between AVs and pedestrians [1, 2, 4], the current project KIRa focused on the cooperative interaction between AVs and cyclists. In both projects, researchers in traffic psychology and communications engineering jointly explored relevant questions regarding the analysis of human behavior in road traffic and the development of suitable algorithms. In the following two chapters, the results obtained in KIRa are presented, first from a psychological and then from a communications engineering perspective (see next book chapter by Raeck et al.).

1.1 Space-Sharing Conflicts Between Cyclists and AVs in Low-Speed Areas

In this project, cooperative interaction between cyclists and AVs was primarily investigated in urban low-speed areas such as parking lots or shared spaces. We assume that these low-speed areas are often characterized by (1) a shared infrastructure for vehicles and cyclists (e.g., no dedicated bike lanes), (2) less formal rules about priority (e.g., no traffic light control), (3) a higher probability that a road user will change its current behavior (e.g., starting or stopping) and (4) the (partial) occlusion of road users (e.g., due to parked vehicles).

These characteristics have the potential for space-sharing conflicts between cyclists and AVs. According to Markkula et al. (2020), a space-sharing conflict represents “an observable situation from which it can be reasonably inferred that two or more road users are intending to occupy the same region of space at the same time in the near future” (p. 736) [16]. As space-sharing conflicts between cyclists and vehicles can lead to safety-critical situations, it is necessary to either anticipate and avoid such conflicts (e.g., through recognizing the intentions of cyclists) or to handle and solve them safely (e.g., using appropriate communication cues).

1.2 Recognizing Intentions of Cyclists in Low-Speed Areas

To enable an AV to anticipate space-sharing conflicts, it can be useful to recognize the intentions of cyclists. Trajectory prediction can be used to determine the cyclist’s next trajectory using the current state. Thus, potential conflicts with the AV’s trajectories can be identified. In contrast, the starting of a cyclist, as a typical scenario especially in low-speed areas, cannot be determined well using trajectory prediction. Therefore, further information needs to be included. Previous research with pedestrians showed that human observers can use the body posture of pedestrians to predict their intention to cross the street, even when certain information of the body posture (e.g., head or legs) is occluded [23]. However, due to a lack of research, it is unclear whether the body posture of cyclists can contribute to the recognition of cyclists’ intentions. Therefore, this project aimed to investigate how human observers use the body posture of cyclists to predict their intention to start. Further, because body parts of cyclists can be occluded in low-speed areas, it was also aimed to examine the prediction accuracy in these scenarios.

1.3 Communication Between Automated Vehicles and Cyclists in Low-Speed Areas

Implicit and explicit communication can help to solve space-sharing conflicts between AVs and cyclists safely and efficiently [10]. Implicit communication cues refer to the behavior of road users that, on the one hand, change their movement (e.g., vehicle deceleration) or perception (e.g., head turning), and, on the other hand, can be used by other road users, for example, as a sign of the willingness of a pedestrian to cross the road [16]. Explicit communication includes signals with no effect on one’s own movement or perception, such as light signals to indicate intentions of an AV [16].

Several studies, however, have shown that implicit communication cues are used more frequently for interactions between vehicles and pedestrians in low-speed areas compared to explicit communication cues [7, 14]. From a reanalysis of a naturalistic cycling study, it can be assumed that priority between cyclists and vehicles in low-speed areas is similarly more likely to be negotiated using implicit communication cues [3]. For example, agents (i.e., vehicles or cyclists), who reach the conflict space earlier, often take the chance to solve the space-sharing conflict through accelerating or avoiding behavior [3]. Therefore, the present project focused on implicit communication and, in particular, investigated cyclists’ gap acceptance and vehicles’ deceleration maneuvers. It was examined how different time gaps, the vehicle size and speed affect the decision of cyclists to cross a street in front of a vehicle. In addition, differences in the perceptual decision-making process involved in the detection of vehicle deceleration by VRUs were further analyzed using a drift-diffusion model. The results can provide important implications for a situation-specific parameterization of vehicle deceleration maneuvers. Moreover, the results can indicate when explicit communication cues are necessary to support the decision-making process.

1.4 Investigating Space-Sharing Conflicts Between Automated Vehicles and Cyclists

In the previous project KIVI, video recordings from the perspective of a pedestrian at the curb were used to investigate pedestrians’ gap acceptance and deceleration detection performance [1, 2, 4]. However, this methodology had to be adapted to the perspective of a cyclist. Riding with a camera on the bicycle handlebars often resulted in blurry recordings due to the pedaling activity. In addition, in such real-world recordings, it was often difficult for the cyclist to keep a constant speed, on the one hand, and to keep the time gap to a moving vehicle, on the other hand. Therefore, this project did not use real-world recordings, but rather a VR cycling simulation which will be presented.

1.5 Aims of the Research Project “KIRa”

The following sections will provide a rough overview of the research within the project “KIRa” (Cooperative interaction with cyclists in automated driving) regarding the research topics described above:

  1. 1.

    Investigating the body posture as a predictor for the starting process of cyclists.

  2. 2.

    Development and validation of a VR cycling simulation.

  3. 3.

    Experimental evaluation of a drift-diffusion model for vehicle deceleration detection.

  4. 4.

    Investigation of factors influencing the gap acceptance of cyclists.

This chapter aims to give an overall summary of the project activities. Detailed information on the experiments can be found in the related publications at the end of each section.

2 Investigating the Body Posture as a Predictor for the Starting Process of Cyclists

A typical scenario in urban traffic, especially in shared spaces and parking lots, is the starting of a cyclist. Here, starting is understood as the process between getting on the bike and the final roll-off. It is assumed that recognizing the progress within the starting process can be crucial for an AV to decide whether it still takes priority (in early stages of the starting process) or rather yields priority (in later stages). Therefore, this project task intented to investigate how accurately the progress of the starting process can be detected based on the body posture of cyclists. Furthermore, we investigated the importance of different body parts in this rating as well as the accuracy of the ratings when body parts are masked. It is assumed that occluded parts of the cyclists’ body, e.g., due to signs, parked vehicles or the bicycle frame, could be a highly relevant problem in shared spaces. The results could support the development of efficient algorithms for intention recognition and, based on human abilities, allow conclusions about how good algorithms’ intention recognition needs to be at a minimum.

For the examinations, 12 cyclists were recorded while starting with their bicycle. The recordings were taken on a parking lot from two different perspectives of a car driver: The view from behind and from the side. The period of interest was the time between getting on the bike and the final roll-off. The recordings were split into four images per second to allow a better visibility of the cyclists’ body posture. Further, these images were used either without masking or (after image manipulation) with masked upper or lower body. During six standardized experiments, these images were presented both in chronological (baseline condition) and random order (experimental condition). In the chronological order, participants were able to build up prior knowledge about the progress based on the images before. In the conditions with random order, prior knowledge was not available and the ratings were possible based on body posture only. For each image, participants were asked to provide ratings about the progress in the starting process using a scale between 0 and 100%. Further, the participants were asked to specify which parts of the cyclist’s body were relevant to these ratings.

Surprisingly, the ratings of progress in the starting process were similar between the baseline and experimental conditions. In the conditions with random order, the ratings often increased in the order the images were originally taken. Thus, the randomly presented images of a cyclist could be rearranged well into chronological order. This could be shown for the different perspectives as well as the masked images. In particular, the ratings at the end of the starting process, shortly before the cyclist accelerated, were rated very accurately in each of these conditions. Furthermore, a lower variance could be observed in the ratings at the beginning and at the end of the starting process. It is assumed that the beginning and the end of the starting process are associated with characteristic body postures. Regarding the relevant body parts, the legs showed the highest ratings as the decisive part at the beginning. As the starting process continued, the importance of the legs decreased, while the importance of the upper body, head and feet increased. When body parts were masked, the remaining parts were substantially able to compensate the occlusion.

The results showed that the progress in the starting process can be recognized accurately based on body posture even when body parts are masked. Thus, it seems possible that the final roll-off can be detected early and ensures a safe interaction. Further analysis is required to formally describe the body posture in the starting process in order to implement these characteristics in algorithms for the intention recognition.

Related Publications:

  • [30] Trommler, D., Ackermann, C., Krems, J.F.: Investigating the body posture as a predictor for the starting progress of cyclists. In: 33rd International Co-operation on Theories and Concepts in Traffic safety (ICTCT) Conference, Berlin, Germany (2021).

  • Trommler, D., Krems, J.F.: Using cyclists’ body posture to support a cooperative interaction in automated driving (in prep.).

3 Development and Validation of a VR Cycling Simulation

3.1 Development of a VR Cycling Simulation

While different virtual reality (VR) cycling simulations are available in the entertainment and sports sectors, these commercial products are of limited use for research purposes. For the examination of the interaction between AVs and cyclists in this project, a VR cycling simulation has to meet several requirements: (1) Accurate control of road user maneuvers, including their trajectories, speed and speed adaptations. (2) Realistic physical visualization of the environment and vehicles, including gravity and other physical forces of objects. (3) Detailed data recording. (4) Ideally, a user-friendly graphical user interface (GUI). (5) Cost-efficient development, especially regarding the hardware requirements.

For this purpose, different existing VR (driving) simulations were compared. These include CARLA [8], OpenDS [17], VICOM Editor (TÜV DEKRA arge tp 21), STISIM driving simulator platform, and Westdrive & LoopAR [19]. Considering the criteria mentioned above and seeing as it offers the opportunity to modify the VR driving simulation to a VR cycling simulation, Westdrive & LoopAR [19] was chosen. This VR implementation is based on the Unity 3D game engine. Therefore, a realistic physical behavior of all objects is ensured by the Unity3D physics engine. Due to the open-source implementation and the availability of a GUI, an individual, simple and fine-grained design of road user maneuvers is possible. Likewise, the data recording can be accurately adapted to the individual research questions. Lastly, the hardware requirements reflect the specification of a modern desktop computer.

However, Westdrive & LoopAR was originally developed for studies on automated driving from the passenger’s perspective, specifically to investigate takeover requests [19]. To generate a naturalistic impression of a bicycle ride, the VR simulation was adapted to the cyclist’s perspective, showing the moving bike, the handlebars and the cyclist’s hands in the foreground. The VR cycling simulation can be displayed on three different monitors to provide the view to the front, right and left. In laboratory studies, these three monitors can be placed in front of a static bicycle on which participants can sit. When implemented as an online study, the VR cycling scenarios can be saved as video files with a frontal perspective.

After these adaptations, VR cycling scenarios can be implemented at relatively low cost and the setup seems to be suitable to investigate communication signals of AVs interacting with cyclists in a safe and replicable way. A disadvantage of this implementation is that the participants cannot control the behavior of the cyclist in the VR and thus, behavior of cyclists interacting with AVs, such as braking or avoiding the vehicle, cannot be studied directly.

3.2 Validation of the VR Cycling Simulation in Terms of Perceived Criticality as Well as Experience of Presence

Several validation studies were conducted using the VR cycling simulation, for example regarding the perceived criticality and experience of presence. It was aimed to investigate whether space-sharing conflicts between cyclists and vehicles with varying proximity are associated with the perceived criticality. Three typical scenarios were evaluated: (1) A vehicle exiting a parking lot and crossing the bike lane in front of the cyclist. (2) An intersection with a vehicle approaching from the left and crossing in front of the cyclist. (3) And a vehicle turning to the right and crossing the bike lane in front of the cyclist. The criticality within each space-sharing conflict was varied using the initially attempted post encroachment time (IAPT) [6]. The IAPT is defined as the time interval between one road user leaving a conflict point and another road user entering the same point, assuming no behavioral changes, such as speed changes, are initiated. Lower IAPT values are associated with a closer proximity between the two road users and thus with a higher potential for a critical outcome of the space-sharing conflict. In this validation study, the IAPT values ranged from one to three seconds for each scenario. Additionally, a baseline ride was performed for each scenario in which the crossing vehicle was absent. The perceived criticality was assessed using a scale developed by Stange et al. [26]. In addition, the experience of presence in the VR cycling simulation was evaluated using the Igroup Presence Questionnaire [24].

An online study was conducted with N = 35 participants. Each scenario with each IAPT level (including the baseline trial) was presented twice with a subsequent questionnaire on perceived criticality. At the end of the study, the experience of presence was assessed. The analysis of the perceived criticality revealed that the baseline rides were rated as significantly less critical compared to the rides with a space-sharing conflict (except for the turning scenario, which showed only a significant increase of the perceived criticality for IAPT = 1 s). Furthermore, the conditions with lower IAPT values were rated as more critical in each scenario. In addition, the results revealed that the turning scenario was perceived to be more critical compared to the intersection and parking scenarios. The analysis of the experience of presence indicated an acceptable experience of presence with a moderate score in the general presence dimension and a good score in the spatial presence dimension.

Based on these results, it is assumed that the VR cycling simulation is suitable to investigate cyclists’ perceived criticality in interactions with AVs. Therefore, the simulation can support the development of safe and comfortable driving maneuvers of AVs in space-sharing conflicts. Further, an acceptable experience of presence in this VR cycling simulation can be assumed for online studies. It may be expected that the experience of presence will further increase when the cycling simulation is used in laboratory studies with a static bicycle in front of three monitors or using a VR headset.

Related Publication:

[31] Trommler, D., Bengler, P., Schmidt, H., Thirunavukkarasu, A., Krems, J.F.: Validation of a VR cycling simulation in terms of perceived criticality and experience of presence. In: Petzoldt, T., Gerike, R., Anke, J., Ringhand, M., Schröter, B. (eds.) Contributions to the 10th International Cycling Safety Conference, pp. 235–237, Dresden, Germany (2022). https://www.icsc2022.com/wp-content/uploads/icsc2022_book_of_abstracts.pdf.

4 Experimental Evaluation of a Drift-Diffusion Model for Vehicle Deceleration Detection

Vehicle deceleration can be used as an implicit communication signal to give priority to VRUs [34]. Results of the previous project KIVI showed that the detection performance of vehicle deceleration by VRUs may depend on various factors, such as deceleration rate, initial speed, age, and gender [1]. To provide a detailed understanding of the underlying differences in decision-making, these effects were further analyzed using a drift-diffusion model.

According to these models, perceptual decision-making is based on an accumulation of sensory evidence over time until a boundary is reached [22]. Several parameters are used to describe this process, which correspond to different components of the human information processing. The most important parameters are (1) drift rate, which describes the rate of evidence accumulation and is associated with the quality of evidence, (2) boundary height, which is related to the amount of evidence for a decision and reflected by the response caution in decision-making, (3) starting point, which can be positioned closer to the boundary in case of expectations towards a decision and 4) the non-decision time, which summarizes the time interval for stimulus encoding and motor response execution [22]. Using reaction times and response accuracies from empirical experiments, these parameters can be estimated after a model fitting [25].

This project task intented to investigate how deceleration rate and vehicle speed affect the parameters of a drift-diffusion model. A study was conducted with N = 62 participants which saw videos of approaching vehicles that either decelerated or not. These videos were recorded for the previous project KIVI. A detailed description of the video recordings can be found in [1]. The participants were instructed to press keys indicating whether the vehicles decelerated or not. In case of deceleration, the slowing down process was initiated immediately after the video onset. The deceleration rate (−1.5 and −3.5 m/s\(^{2}\)) and the vehicle speed (20 and 40 km/h) were varied as independent variables.

After the model fitting, the results showed substantial differences in the drift rate depending on the deceleration rate. This is consistent with the assumption that a higher stimulus quantity (i.e., higher deceleration rates) leads to faster evidence accumulation. Moreover, the boundary height as a measure of response caution varied slightly between the conditions with low and high vehicle speed. Higher values for the boundary height were observed for the conditions with higher vehicle speed. Additionally, there was a slight increase in non-decision time for the conditions with higher vehicle speed. This suggests that stimulus encoding needs slightly more time in the conditions with higher vehicle speed than in the conditions with lower vehicle speed. Moreover, a slight shift of the starting point towards the decision that the vehicle does not decelerate could be observed in conditions with higher vehicle speed. This finding suggests a decision bias.

In summary, a good model fit to the empirical data was achieved. The results showed that the contextual factors influenced the model parameters in a way that are in line with theoretical considerations. Further studies might investigate whether a complementary use of explicit communication signals, especially for slow decelerating and fast moving vehicles, leads to an improvement of the evidence accumulation process and thus to a higher satisfaction and perceived safety of VRUs in interaction with AVs.

Related Publication:

[29] Trommler, D., Ackermann, C., Krems, J.F.:A drift-diffusion model to explain vehicle deceleration detection of vulnerable road users. In: Stewart, T.C. (ed.) Proceedings of the 19th International Conference on Cognitive Modelling, pp. 289–294 (2021). https://acs.ist.psu.edu/papers/ICCM2021Proceedings.pdf.

5 Investigation of Factors Influencing the Gap Acceptance of Cyclists

When modelling human-like deceleration maneuvers for AVs, the deceleration rate as well as the time of a deceleration onset that VRUs expect for safe crossings need to be considered [4]. This expectation can be investigated through the VRUs’ gap acceptance which defines the (time) gap that is acceptable for crossing in front of a vehicle [4]. However, previous studies on pedestrians’ perspective show that the gap acceptance behavior may depend on external attributes (e.g., vehicle speed, vehicle size and time to arrival) as well as on internal attributes (e.g., gender and age of VRUs) [28]. Building on these findings for pedestrians, this project task aimed to examine the gap acceptance of cyclists. Therefore, the objective was to investigate the effects of vehicle size (car vs. truck), vehicle speed (20 vs. 40 km/h) and different levels of the time to arrival (TTA; ranging from one to five seconds).

The videos were generated in the presented VR cycling simulation with a length of approximately 10 s each. The videos were shown from the perspective of a cyclist riding towards an intersection while a vehicle is approaching from the left. Traffic signs and the study instructions indicated that the cyclist does not have priority. The TTA was measured as the time gap to the vehicle when the cyclist reaches the (theoretical) collision point at the intersection. N = 35 Participants were instructed to indicate by pressing a key, whether or not they would cross the road in front of the vehicle. As dependent variables, the crossing decision and the time of this decision before reaching the (theoretical) collision point were recorded.

The results revealed that more participants decided to cross in front of the vehicle as the TTA level increased. Further, for each TTA level, the willingness to cross was higher in conditions with faster vehicles than in conditions with slower vehicles. Within the majority of conditions, slightly more participants chose to cross in front of a truck compared to a car. Regarding the decision time, the results showed that the decision to cross or not is made approximately between two to four seconds before reaching the intersection. In conditions with faster vehicles, participants decided later (i.e., the cyclist was closer to the intersection) than in conditions with slower vehicles. In contrast, the decision was made earlier in conditions with lower TTA levels compared to conditions with higher TTA levels. Similarly, the decision was made earlier in conditions with trucks than in conditions with cars.

A further analysis focused on differences between the age of participants. For this, the sample was divided into two groups with 18 younger (18–35 years old) and 17 older (>35 years old) participants. The results indicated that participants’ crossing decisions were similar, with the exception of the condition with 5 s TTA, where more younger participants than older ones expressed their intention to cross. Furthermore, older participants tended to make their crossing decisions later than younger participants.

To sum up, similar to the results for pedestrians, it is assumed that no universal parameterization is possible to design informal communication between cyclists and AVs. The study revealed that there are substantial differences in the gap acceptance of cyclists depending on vehicle size, vehicle speed and TTA. The findings also suggest the importance of considering age as a factor. Further, the results showed differences in decision time. The decision not to cross is made earlier than the decision to cross in front the vehicle. Therefore, AVs should use communication signals for giving priority early, especially when the TTA level is low and/or the AV is a truck. In future studies, additional factors, such as internal (e.g., age of the cyclist) and external attributes (e.g., time of day), need to be explored. Moreover, it could be relevant to investigate the decision-making process using drift-diffusion models as proposed in the previous section.

Related Publications:

  • [32] Trommler, D., Springer-Teumer, S., Krems, J.F.: To ride or not to ride: exploring cyclists’ gap acceptance in the interaction with (automated) vehicles. In: 34th International Co-operation on Theories and Concepts in Traffic Safety (ICTCT) Conference, Györ, Hungary (2022).

  • Springer-Teumer, S., Trommler, D., Krems, J.F.: How do vehicle size, speed, TTC, age and sex affect cyclists’ gap acceptance when interacting with (automated) vehicles? In: 1st International Conference on Hybrid Societies, Chemnitz, Germany (2023).

6 Summary

This chapter gave an overview of the research project “KIRa”, which investigated the cooperative design of the interaction between automated vehicles and cyclists according to four project aims.

First, the body posture was investigated as a predictor of the cyclists’ starting process. The results showed that the progress of the starting process can be accurately detected based on body posture by human observers. This was even possible with a high accuracy when certain body parts (e.g., head or legs) were masked. Thus, it seems possible that an AV can recognize a cyclist’s intention to start and can either avoid safety-critical situations with cyclists or can resolve them cooperatively at an early stage.

Second, the development of a VR cycling simulation was presented, including its validation in terms of perceived criticality and experience of presence. The findings revealed that the VR cycling simulation is suitable to investigate the cyclists’ criticality perception when interacting with AVs. Different levels of proximity between a vehicle and a cyclist in three different shared-space conflicts reliably resulted in corresponding changes in the perceived criticality. This was investigated for different scenarios. Therefore, it is assumed that it is possible to investigate maneuvers of AVs interacting with cyclists in a standardized, reproducible and safe way.

Third, a drift-diffusion model for vehicle deceleration detection was empirically validated. The model parameters suggested the applicability of drift-diffusion models in applied research areas such as automated driving. This can lead to an improved understanding of the decision-making process of cyclists and further to the design of implicit and explicit communication signals adapted to humans’ information processing abilities. This is expected to increase cyclists’ acceptance and trust towards AVs.

And fourth, factors influencing cyclists’ gap acceptance were investigated. The effects found for the gap acceptance of pedestrians could be confirmed, such as a strong dependence of the gap acceptance on the time gap, the vehicle size and vehicle speed. Furthermore, the same factors were associated with different decision time of cyclists (i.e., whether they would cross in front of the vehicle or not). These results highlight that cooperative interaction between AVs and cyclists is closely linked to context-sensitive communication.

7 Further Reading

In addition to the publications in this project, we will refer to the publications of the previous project “KIVI”, which investigated the interaction between pedestrians and automated vehicles: