Realistic and comprehensive traffic simulations are essential for the effective testing and evaluation of emerging technologies, such as Vehicle-to-X (V2X) communication, and diverse use cases, particularly within complex urban environments. While current traffic scenarios often focus on motorized vehicles, there is a need to address the safety of vulnerable road users (VRUs), such as pedestrians and cyclists. This is especially relevant in light of the European Union’s Vision Zero initiative, which aims for zero road fatalities by 2050. Although a few pedestrian-focused scenarios exist, there is no scenario specifically addressing bicycle traffic, despite their status as one of the most at-risk VRUs, with stagnant fatality rates in recent years. To address this gap, this paper introduces the Hanover Traffic Scenario for SUMO (HaTS), a novel traffic scenario including motorized vehicles and bicycles. HaTS provides a detailed and accurate representation of the road network, traffic light systems, and buildings within the city center of Hanover, Germany. A key feature of HaTS is its integration of real-world traffic count data for both bicycles and motorized vehicles, enabling a realistic and representative traffic demand representation. Additionally, a novel metric is employed for the parametrisation of the scenario, enhancing the alignment between real and simulated traffic volumes. For the validation we compare the results of the HaTS with the real world traffic counts. HaTS is the first open-source SUMO scenario focused on bicycles, providing a realistic representation of the road network and traffic demand, thereby contributing to the advancement of urban traffic simulations.
—
Title: HaTS – Hanover Traffic Scenario for SUMO
Presenter: Nico Ostendorf
Authors: Nico Ostendorf, Keno Garlichs and Lars C. Wolf
[Music] an over traffic scenario for sumo and first first of all the little motivation why we need a new scenario because there are already a lot of scenarios but yes I don’t think I need to tell you but simulations are essential to test various application and technologies especially for use cases which are relevant for safety or if testing is just quite hard in real Examples would be vehicle-to-X communication or for example traffic forecasts and simulations currently mainly focus on motorized vehicles or rails and so on trains and but there’s a growing need to include also so-called vulnerable road users in simulations. So vulnerable road users include for example pedestrians which are already well implemented in sumo I think and also bicycles and if you can look at the figure one you can see a model split and the model split forecast. So the first part is the model split of the year 2022 and you can see that um there are already a high amount of VR use inside but if we now look at the model split forecast worldwide for VUS then you see that the amount of V use increasing in the next years and we have a so-called vision zero target in the EU and this tries to eliminate all road fertilities under the year 2015. 50. And one big problem here is that we have in the last years a stagnant in fertility rate for bicyclist and therefore it’s very important to also include bicycles now in simulations because they could profit yes very highly from new technologies especially from safety technologies. And right now we didn’t find any yes realistic scenario with real world data which really includes a high bicycle density for Zumo and also for other simulators. There was no such uh scenario and therefore our goal was to create such a scenario with a high bicycle density but not only to include bicycles but also have vehicles inside of the scenario. And all of these should uh be based on real world traffic counts because yes for our use cases it’s important to have a yes realistic representation of the city. Um for the city we choose as the title mentioned already. Um we choose the city because it has one of the largest proportions of bicycles in the model split uh in cities with more than 500,000 inhibits. Antonova also um provides traffic counts for bicycles and vehicles which is not uh very normal in German cities. So first of all how do we want to approach our goal? So in the first step we selected and generated the map and the road network. In the next step we did a analysis of the traffic counts we got from the city of Anova. And then we did the parameterization of our scenario. And in the last step we uh validated our handover traffic scenario. So first of all the selection and generation of the map and road network. So in figure two you could see the uh yes the region of Anova we choose for our scenario. So it’s not the whole city of Anova, just a short part of Anova, but this part has a very high traffic volume compared to other locations in Hanover and also a very unique topology inside of Anova because it includes various primary streets which are highly used during the morning and evening hours. Then we have inside of this residential neighborhoods and it’s also located near to the city center. And all of these guarantee that there’s always or mostly a very high amount of uh vehicles inside of the scenario at this part. But there’s also this waterfront you can see there. And at this this waterfront is a very focal point for cyclists not only to go to their work and back because it’s very pretty there but also to uh go for a bicycle ride or something like this. And therefore it guarantees also a very high amount of bicyclist in Hanova. So in the picture on the left side you could also see these blue uh triangles and these are our traffic counts we got. So this are seven distinct traffic counts available mostly located on the main street sadly but yes so just some network statistics about our simulated networks then the total area is around 5.5 square kilometers um we have a high amount of exclusive vehicle and bicycle roads uh yes the bicycle roads are doubled the size of the vehicles because at the most roads there are two bicycle lanes if there’s one road. Um, and there are also roads where both vehicles are allowed and the scenario includes about 42 traffic lights. Um, yes, now I will continue how we get this part of handover inside of the sumo simulator. So, we did it as I think the most of us would do. Uh, we use the OSM web result. But one big problem with the OSM or with OSM map data is that the crossings and speeds are not very well represented inside of OSM. And if you now look at figure three, I don’t know if you could see it very good here at figure 3A. Um, vehicles and bicycles were allowed at this crossing to go in any direction. And also the representation of the crossing itself was not very perfect. And if you would simulate it like this, then it would be a quiet mess. And then we looked into Google Maps and and because I’m from Anova, I also had some knowledge about this crossings. Um, and then we also looked into Google Street View and used all of this to make the manual correction of all the crossing inside of the scenario, all the streets and also the placement of the traffic lights and also the buildings. Um and also we removed the unused areas because to simplify the procedure because it would take a lot of time to also correct those. So therefore there are no pedestrian ways ra ways uh private streets or bus lanes inside of our scenario. And on figure four you could see the final result of our scenario. Um so in the next step we wanted to have traffic inside of our scenario. So therefore we it was necessary to analyze the traffic counts and do also the parameterization of the scenario. So, first of all, the traffic counts. As I told you before, we have seven distinct traffic counts at distinct crossings, and they were all all collected in May 2022 from uh 6:00 a.m. to 7:00 p.m. And the data is a 50 minutes resolution. Um, and each traffic count is described by a so-called turn count for vehicles and for bicycles. So, an example of such a turn count you could see in figure five above. Um so a turn count says which vehicle or how many vehicle go from one lane to another lane and therefore you have a really good resolution of one crossing inside of your traffic counts. For example here 44 vehicles went from the southern side to the western side. And so this all of this leads to 86 distinct turn counts. We had for the simulation or for the parameterization and trip generation. And we did this with the yes sumo tools. So the root sampler and the random strip uh drip script and in figure six you could see the traces inside of the traffic counts. So yes in blue there are the vehicles in orange the bicycles as assumed there are more vehicles than bicycles available. But um you can see this pattern that in the morning there’s a peak and also in the evening then in the middle of the day it’s a little bit lower and in the end of the day and beginning also and it’s nearly the same for bicycles but here’s a short difference because in the end of the day yes a lot of people will go on the bike to make a good trip or something like this. So after we generated our um trip uh roots inside of the sumo scenario, we first of all parameterized the scenario. Um so the start and the end value are representative for 6 a.m. and 7 p.m. exactly like in the traffic count data. Our default step length is uh way shorter than the default value. So we selected a value of 0.1 seconds, but we also tested shorter and higher values. So we selected this value because our use case for this scenario is V2X communication and in V2X communication a resolution of 0.1 seconds is essential because of the generation rules for messages. Um also we activated the sublane model for this scenario because as most of you may know bicycle will never wait uh behind each other at a crossing. they will always break beneath each other also overtake each other in a single lane and so on. And we also wanted to minimize the impact of junction blocking vehicles. So in real world you would also see things like you try to avoid a vehicle which is on the junction and no one will wait when he has a green light. And then also a very important parameter was a car falling model. So for the vehicles we choose the default uh cow model but for bicycles uh we we presented last week our realistic bicycle dynamics model. So this is able to make a more realistic longitudinal behavior of bicycles inside of the simulator. So more realistic acceleration and deceleration values and also speed values. And we used also the Krauss model. So it’s not only tested on our new model. Um then you may already saw there was another parameter called the rerouting probability. Um this parameter allows you to reroute your planned route if there is a jam on your route or if the if it takes a shorter time if you choose a new route. Um and this is yes behavior like uh realtime navigation on Google maps and without using this parameter you can see one result here down for the vehicles. Um the whole scenario was jamming because there were some crossings where it was necessary to avoid the vehicles there and need they needed a rerooting and if the time of the day continues there was so much dam scenario that it was not possible that any vehicle was turning anymore and yes to optimize such a parameter it’s an optimization problem and we choose an objective function for this um we used a combination of three different uh metrics which are all weighted equally and also bicycles and vehicles was weighted equally afterwards. Uh first of all we choose the normalized root mean square error for this. Uh we selected this one because um related work like the in scenario also uh used this one to um find the perfect rerooting probability. But when we used only this value, we had this problem that the peaks of the day were represented very bad. The whole day was okay, but in the peak times it was bad. And especially for our use case of V2X communication, it’s important to have also the peaks. Therefore, we also uh use the absolute traces difference. So the uh difference of the whole day between simulated and expected traces and the maximum time slot difference because we also wanted that at each time slot inside of our simulation it’s as perfect as possible and therefore we selected the highest difference at one day at one time slot here. And as you can see now on the figure eight these are the results of our scoring. Um so for low values uh we could not reach the perfect traffic amount but also for very high values we could not match the traffic amount of the traffic counts because then most vehicle will reboot and then you will see that they are at other crossings and so on and that’s not what we wanted. So but there was a quite good uh field or range of values and we selected a value of 0 uh 22 equal to a probability of rerooting to 22%. So in the last step we did the validation of our scenario. So we compared again with our real world data and in the picture above you can again see the comparison between simulated and expected traces. So in orange the expected values from the real world traffic count data and in blue the simulated amount of traces. Uh in figure nine it’s for the vehicles. In figure 10 it’s for the bicycles. And as you can see it’s fits quite well to the distribution. We have always a little bit less traffic compared to the real world. One reason for this is the impact from bicycles to vehicles. For bicycles, it’s a little bit higher compared to the vehicles because bicyclists yes tend to break rules at crossings and it’s yes not possible in simulation to have a high such as high amount as in real world which cross the green light. So therefore the amount of bicycles is always a little bit higher or sometimes for the peaks also delayed a little bit. Um but yes, we wanted to be honest and we don’t have these only these best cases and not every crossing was able to represent perfectly. So we uh looked at each turn count inside of our simulation and in the real world on com and compared them with the expected and simulated values. Of course, most of the crossings are represented like the best case and we have a very good distribution compared between uh simulated and expected values, but there are also some worst cases where it was not possible to recreate the amount of traffic inside of the simulation. So, for example, we have here one crossing or turn count. This was a left turn without a green or explicit green light. And here it was very hard in simulation to have the same amount of vehicles going left as in the real world. So in this case there was often a rerouting because there was a jam on at this crossing. Um now also yes some of the limitations of our scenario. So um yes as you see in the first picture of the location of the traffic counts we have a very geographic concentration of this traffic counts inside of our road network. So we want to have more data but there’s no data available sadly and therefore um yes without this data it was not really possible to also validate the residential areas of the traffic amount and also validate the patterns in all the locations we wanted to simulate. Um then there’s also yes an unknown part. So we got the traffic counts from the city of Anova. They told us how they collected them. They the collection was b based on video analysis. And as most of you may know video analysis is always very sensitive to weather, light conditions, the technology which is used to and yes to get the number of the vehicles inside. And also if you now assume there’s a big vehicle in front of bicycle, you will never see this bicycle at this um video. Um and therefore they could not report us any error margin inside of these traffic counts. So we have to yeah trust them that the data is accurate enough. Um then there was also no yes dynamic traffic patterns inside of our data. So we could not evaluate if the traffic flow itself is really accurate. So there was no data about the average speeds of vehicles or how long a vehicle would take from one point to another inside of or in the real world. So therefore it was sadly not possible to also evaluate the things. So last but not least a short summary. Um the first thing is that we have now a realistic representation of the road network and ANOVA able in simulations and also in sumo. And then the next summary is we have also a realistic or as realistic as possible representation of bicycle traffic and vehicle traffic in Hanover now in simulators. And the last one is more like an information. Uh all of this what we have done is open source available. You can see it under the GitHub link or just scan the QR code. Yes. Thanks a lot for your participation and if you have any questions.