1 Introduction

During events and disasters, confusions may occur due to the concentration of people, and secondary damage may also occur due to accidents or failure to complete evacuation. It is important to predict whether these problems will occur and to propose solutions.

Since FY2021, in collaboration with RIKEN, Kobe City, and NTT DOCOMO, Inc. (hereafter DOCOMO), we have made an attempt to construct a simulator (i.e., a digital twin [1]) that can reproduce realistic and practical situations by introducing digital maps and usual human flows during events or disasters, which are based on real data for Kobe City.

The map used in this simulation was generated from Open Street Map [2], hereafter OSM. Since there were many discrepancies between the data and reality, it was necessary to examine the aspects that needed to be corrected and to modify the maps to make them usable. The details of the map and the results of a simulation using the map will be demonstrated.

2 simulation method

The core elements of this research method are; simulation framework, maps, a realistic origin destination pairs (hereafter OD) based on real population data.

2.1 Simulation framework

To implement the simulation, we used CrowdWalk [3], hereafter CW, an agent-based open-source pedestrian simulator. To run the simulation, we need to provide a digital map, origin and destination link/nodes for each agent, and intermediate paths if desired. The shortest path for each agents are automatically generated when the simulation starts then the walk is initiated. If the intermediate paths are specified, shortest paths that include them in series will be used.

The interaction between pedestrians is calculated by the social force model, which automatically generates congestion or free walking conditions [4]. The scripts are equipped with the ability to convert and import maps from Shape Files or OSM.

As a candidate for usable map data, maps that include pedestrian walkable streets, pedestrian signals, and connection information are rare, both commercially available and free. In the specific area of Kobe City shown in Fig. 1 treated in this study, no such map could be found.

Fig. 1
figure 1

Simulation and analysis region in the present work, shown as the region enclosed in red

Full size image

In this study, a pedestrian map was created by importing OSM data into CW and then manually processing it. The following methods were used to implement signal control.

2.2 Preparation of the pedestrian map

As shown in Fig. 2, sidewalks are partially implemented in the OSM, but they are not implemented in the majority of the areas of this simulation. However, when importing the OSM into CW, it is possible to convert the roadway into a walkable pathway for pedestrians, and this feature was used to ensure connectivity.

Fig. 2
figure 2

Partially implemented pedestrian walkway. The left half of the map is not implemented with them, while the right half is

Full size image

The above method is not very realistic for roads with sidewalks implemented on both sides of the vehicular road, as all three roads would be walkable. It affects the total road width available for pedestrians. Initially, we thought that this artifact could be avoided by eliminating all sidewalks if we could assume that connectivity was ensured by the vehicular road, but it turned out that there are places implemented in OSM that can only be reached by sidewalks, and connectivity to the corresponding locations would be compromised. For this reason, we decided to leave both sidewalks and roadways as are.

The pedestrian signals affect pedestrian flow and act as a congestion factor, so they need to be replicated. By specifying the |CloseGate| process in the CW scenario file for all roads (links) with the |Crossing| tag, pedestrian movement on or entering the corresponding crosswalk is prohibited, thus the signal behavior is reproduced.

Performing the above is not enough to implement a fully operational signal. There are cases that pedestrians pass through the vehicular way while the signal is red since the signal operates on links. Such persons will pass through the intersection as shown in Fig. 3 left. For this reason, we removed the nodes and links inside all the intersection as shown in Fig. 3 right. By this modification, any pedestrians are blocked to enter or pass the crossing, which is the expected behavior of the pedestrian signals. Note that this implementation is only available for vehicular roads with sidewalks implemented in OSM. The area where the signal is mainly implemented roughly coincides with the most heavily trafficked area of Sannomiya South, as shown in Fig. 4. Signals are not implemented in other areas where sidewalks and crosswalks are not originally implemented. These streets generally coincide with local areas where often no signals are installed in reality, and for the purpose of the simulation, they are not considered to have a significant impact, so they are not implemented at this stage.

Fig. 3
figure 3

Left: schematic diagram explaining penetration of pedestrians with default OSM map. Right: blocking pedestrian with the deletion of roads inside an intersection. In both figures, green and red links represent the walkway and red-signaled crossover, respectively

Full size image
Fig. 4
figure 4

Region signals are mostly implemented

Full size image

Road widths also affect on pedestrian congestion. Although it is possible to introduce OSM data when importing, there is no correct specification in the original data and the imported data is uniformly set at 2 m. However, according to the Japan Road Structure Ordinance [5], it is legally required that sidewalks should be at least 3 m wide to allow pedestrians to pass by each other. In this study, the road widths were set manually by surveying the sidewalk widths of major roads based on Google Map satellite images and Street View, where most pedestrians are expected to take or where they are going to be guided by policies.

2.3 DOCOMO mesh population

To reproduce background traffic, we attempted to use available human flow data. DOCOMO collects GPS location data from mobile phones to record how many people were present in a certain area at a certain time, and sells the relevant data in the form of statistics that are converted so that no individuals can be identified. In this study, the statistical data were provided by DOCOMO as a joint research institute, and were used in the analysis.

The time resolution of the data is 1 h, and the spatial resolution can be selected between 1 km (so-called 3rd mesh) and 500 m (4th mesh). In this case, since the area to be treated is about 3 km from north to south, a 1 km mesh would have resulted in a very coarse 3 \(\times\) 4 mesh. It shall be an insufficient resolution, which could have resulted in an inaccurate spatial arrangement of congestion, so data with a 500 m mesh were used.

As an example of the data, a time variation of the population in the area bordered blue rectangle in Fig. 1 where the mesh including JR Sannomiya Station is shown in Fig. 5. Fig. 6 represents the spatial population distribution during 14:00–15:00 01 November 2021.

Fig. 5
figure 5

Population time variation in the mesh containing JR Sannomiya station

Full size image
Fig. 6
figure 6

Spacial population distribution in Kobe city center during 14:00–15:00 2021/11/01

Full size image

The number of OD pairs as background traffic under normal conditions was estimated from the population increase to the mesh of interest, assuming only from adjacent meshes. N persons increase in 1 h assumes that the population came equally from eight neighboring meshes. In other words, N/8 people from the left, N/8 people from the north, etc., and so on. In the case of a decrease in population, such as at night, the OD is assumed to be in the opposite direction.

In addition to the above, by adding the OD from the planned development area to the existing infrastructure such as stations, we attempt to estimate how congestion would occur in a newly developed city in the event of a disaster. They are assumed to be controlled under an assumed evacuation policy, while background pedestrians are not.

3 simulation results

Assuming OD only between adjacent meshes as based on the estimation method explained in the last section, the total background pedestrian population during 14:00–15:00 01 November 2021 was calculated to be 6509.

Here, let us assume that a disaster occurred and all above pedestrians required to escape to some other cities by train. To try out such an evacuation scenario, destinations of all agents are then changed to their nearest station. Each agent chooses the closest one among Motomachi, Hankyu Sannomiya, and JR Sannomiya as the destination based on the shortest path from its origin. Location of each station is shown in Fig. 7. Agents are generated in the above number under uniform time distribution within 30 min right after the simulation start, to reproduce the severe situation that evacuees rushes towards the station; agent generation time should not dominate the evacuation time. To be realistic, all the evacuees takes time (30 min) to be generated, not at the first second of the simulation either. We also assume that the trains have the capability to carry them all, and no congestion occurs at the station: once each agent arrives at one of the stations, it disappears there.

With a signal cycle of 4 minutes evenly divided by green and red, no congestion occurred except for the vicinity of stations and it took 1:35:10 (5710 s) of in-simulation time for all agents to exit the map. All signals are synchronized and turn red or blue at the same time. Although it would be realistic to have reversed signals for directions that differ by 90\(^{\circ }\), this would require calculating the direction of all signals and setting them separately. In the present setup, the transportation capacity of the road is not expected to be much affected by the signal phase setting, since the human flow occurs mostly in the north-south or east-west direction. The agent behavior of being caught by a traffic signal a few blocks ahead in this setup is the same as the realistic situations where the signal phases are opposite in the east-west and north-south directions. More realistically, the signal period and timing may have a unique setting in each location, but there is not enough data provided to reproduce this. Due to these reasons, for simplicity, we have set all signals to blink at the same period and phase in the present study.

Snapshots with green and red signals are shown in Figs. 7 and 8, respectively, location of the stations are denoted in the former. Note that the latter is with red-colored agents which are temporarily trapped by the red signals, while the former is without, since in snapshots, slowed down and stopped pedestrians are displayed yellow and red, respectively. The average trip time with the standard error was obtained to be 1304 ± 9 s, obtained from the trip time distribution shown in Fig. 9.

Fig. 7
figure 7

Snapshot of the simulation only with the background traffic, while green signals

Full size image
Fig. 8
figure 8

Snapshot of the simulation only with the background traffic, while red signals

Full size image
Fig. 9
figure 9

Trip time distribution of the background traffic, with the same condition as Figs. 7 and 8. Each bin has a width of 5 minutes and note that the overall integral corresponds to the number of entire agents present in the simulation

Full size image

In addition to the above background pedestrians, a simulation was conducted under a scenario that 10,000 pedestrians evacuate from redeveloped Shinko area to the Sannomiya area. As with background traffic, destinations were automatically selected.

They all chose Hankyu Sannomiya because all the agents are generated at the same link. They are generated during the first 1 second at once right after the simulation start. The entire simulation ends at 7:08:05 (25,685 s). The snapshot and trip time distribution are shown in Figs. 10 and 11, respectively. Mean trip time was 9020 ± 60 s. The distribution has an oscillating long tail, which is considered to be the effect that evacuating pedestrians are periodically trapped with signals.

Fig. 10
figure 10

Snapshot of the simulation with the background traffic, evacuees from the Shinko area, and signal. Black and red circles represent the Shinko area and congested road segments, respectively

Full size image
Fig. 11
figure 11

Trip time distribution, with the same condition as Fig. 10, with the same bin width as Fig. 9

Full size image

To evaluate a policy to avoid the congestion above, splitting their destinations and guiding the route was also tried out. Specified routes are shown as R3, 4, and 5 in Fig. 12. 3333 pedestrians are assigned to each route. The summation of them is 9999, which is effectively identical to the case without the route split policy above. The snapshot and trip time distribution is shown in Figs. 13 and 14, respectively. The step in the distribution is due to the different destinations, the component lasts till the end of the simulation is from evacuation to JR Sannomiya station.

Simulation time and mean trip time was 4:56:20 (17,780 s) and 5810 ± 40 s, which are 69% and 64% of the case without destination split. This is considered that congestion leaking out north of the Shinko area seen in Fig. 13 was avoided by dividing their routes and destinations. The above simulation results are summarized in Table 1.

Fig. 12
figure 12

Split routes to guide pedestrians from Shinko area, represented by red lines with labels R3, 4, and 5

Full size image
Fig. 13
figure 13

Snapshot of the simulation with destination split in addition to the settings of Fig. 10

Full size image
Fig. 14
figure 14

Trip time distribution, with the same condition as Fig. 13, with the same bin width as Fig. 9

Full size image
Table 1 Time summaries with signals
Full size table

In a situation of disaster such as an earthquake, car traffic should not be allowed, then evacuees could be able to ignore signals for evacuation. The simulation without signals were also conducted. The results are summarized in Table 2.

With destination split, simulation and mean trip time are reduced to 70% in both. By removing signals, background evacuation simulation time was reduced only to 92%, while they are significantly reduced to 52 and 54% with evacuation from Shinko, for the cases with and without destination split, respectively. The percentages of all the simulation time of scenarios with evacuation from the Shinko area are summarized in Table 3.

Table 2 Time summaries without signals
Full size table
Table 3 Simulation time and percentages with and without policies in evacuation from Shinko scenario
Full size table

4 Discussion

In this study, a policy evaluation was conducted to determine the evacuation time could be reduced by dividing the evacuation destination and routes. It is clear that the aided policy effectively avoids congestion. The signal removal was also tried out, and it was also confirmed that effectively and rather drastically reduce the evacuation time.

It also suggests that these two specific policies to be ‘uncorrelated’ or ‘linear’ each other, under a specific condition. A careful look at Table 3 shows that removing the signal reduces simulation time to \(\sim\)50% regardless of the other policy, and splitting the route reduces simulation time to \(\sim\)70%, again regardless of the other, because 37% in the bottom right of the table is close to 36%, which is the multiplication of 52% in the top right and 69% in the bottom left of the Table.

The above is considered to be that both of these two policies are changing the ‘conductance’ of the whole evacuation route by a specific fraction regardless of each other. Note that this conductance would not be something defined to each road segments because it spreads over all used routes in ‘with split’ case and even selected road segments are different in each case. This should be non-trivial and interesting because the number of pedestrians trapped by the signal can be nonlinearly—which actually seems not—affected by route split, and sometimes bottlenecks appears which directly arises the nonlinear response. It is also clear that there is a limitation in the linearity because the removal of the signal reduces the evacuation time only to 90% in the case only with background pedestrians, not 50% like in the evacuation scenario from Shinko area.

The linear relation of policies above was only suggested in one case. Applying other policies may produce other nonlinear results. For example, it is not yet confirmed how the result changes due to the signal cycle, the amount of the background traffic, the origin and destination, or the number of evacuees. We plan to run and analyze many of these cases obtained by parallel execution of a many-case simulation on Fugaku, to clarify the limitation of the law suggested above in the future.

Kobe is surrounded by mountains and the sea from north to south respectively. The evacuation route from Shinko to the station can be relatively short. Other cities without such boundary conditions may have longer evacuation routes and a longer overall timescale for evacuation. It can also be pointed out that the time saved by removing signals and the rough structure of the trip time distribution may be similar, even though the width itself may vary.

In conducting this study, it became clear that pedestrian maps are rare and that it is still challenging to construct a pedestrian simulation in an arbitrary city, or more specifically, a digital twin. However, the implementation of OSM sometimes extends to sidewalks in a major city like Kobe, and the CW function allows the use of roadways as sidewalks. It became clear that it is possible to construct a relatively practical simulation with relatively limited manual modification over OSM data in the present work.