A Simulation Model for Pedestrian Crowd Evacuation Based on Various AI Techniques

Occurring emergencies in big places makes the evacuation operation difficult due to adding confusion, fear and even vagueness and anxiety to mass residents [1]. Constructed environment, evacuees’ communication, and the plans and measures in place to direct the reactions of participants to environmental conditions and characteristics can influence the evacuation system. Therefore, various evacuation methods appear, such as protective, preventive, rescue, and reconstructive evacuations [2]. The problem of evacuation is not solely the physical movement of evacuees; it is multifaceted and related to the physical and social circumstances, such as the high possibility for hazard, great stage of pressure, and inadequate data. These circumstances illustrate robust communication between environment, danger, egress process, population demographics, and participant behavior [3]. A crowd is a communication of a set of people [4]. Various features belong to crowds; therefore, during an emergency, a number of behaviors would be exposed in community areas [5], and simulation is a means of expecting behaviors via responding to the "what-if" situations [6]. Hence, Crowd evacuation simulation is a method for mimicking the flow of participants for the same goal of evacuation from the same place [4]. In the last two decades, rising dedicated attempts via researchers clearly have been seen to explore crowd evacuation during an emergency. Evacuation dynamics have been widely considered to eliminate deaths and injuries during emergency pedestrian incidences [7-13]. A vast number of models is developed to know individual behaviors, such as how individual react a specific condition and how to deal with varied kind of scenarios in the crowd during an emergency; the entire investigations have been there for creating a foundation of enhancement in managing emergencies [14-20]. Modeling is a way of answering difficulties, thus, a newly developed framework after analyzing and controlling via basic pro-test shows the authenticity of its procedures to be mentioned as a model [4]. Existing models can be separated into three main groups: the first one is classical models, the second one is hybridized models, and the third one is a generic model. Each group comprises some models, in the recent past; various applications to investigate crowd evacuation in both normal and emergency situations were done due to these models.

Microscopic is one of the classical models and the details of the individuals and individual behaviors are precisely identified. However, it has a problem in dealing with the large size of participants. In microscopic, cellular automata, lattice gas, social force, agent-based, game theory, and experimental approaches were designed. In the last two decades, cellular automata models have been created to consider an evacuating group of individuals under different circumstances. These models can be categorized into two groups. The first depends on the associations among situations and pedestrians. For example, in 2007, Varas et al. utilized a two-dimensional cellular automaton model to imitate the process of leaving from a single and double door classroom with complete ability. In this study, to each grid, the structure of the room, obstructions spreading, floor field were measured. Moreover, the effect of panic as a counted dimension was calculated which % 5 possibilities of not moving. The model applied random selection to cope with the collision issue. Therefore, the proposed model changed into non-deterministic via these characteristics. From the simulation result, it was clearly observed that the best locations of the door and evacuation efficiency were not enhanced by substituting a double door with two distinct doors. Finally, for the evacuation time, a number of persons and way out width were considered due to suggesting simple scaling law [21]. 

In 2011, Alizadeh proposed a CA model to analyze process of evacuation in a room contained obstructions and had different configuration of the room, such as exit and obstruction locations, width of the exit, light of the area, evacuee’s psychology, and spreading of the crowd, which had a significant impact on the evacuation process. The case study of this model was the restaurant and a classroom. The impact of the distribution of the evacuees, location, and width of the door on time of the evacuation argued and output of the model was com-pared with some static models [22]. In 2014, Guo, Ren-Yong created a model based on CA with better separation of the area and advanced speed of walking to demonstrate moving away of pedestrians from a room with a single exit door. When experiments were simulated, near the exit they noticed the shape of the crowd was affected by both the separation of the area and advanced speed of walking, duration of the individuals at various positions, and efficiency of the evacuees expressed via two-time indicators. Besides, the association between width and flow of the exit was expressed via this model [23].

In 2015, Li and Han recommended a model for simulating pedestrian evacuation based on extended cellular automata to configure different behavioral propensities in people. Familiarity and aggressive were the two chosen behavioral propensities to be examined via this model. During the experimentations of this simulation, behavioral parameters and pedestrian flow arrangements were verified. The output of this study demonstrated that evacuation time de-creases with increasing individuals’ familiarity and increases with the increasing individuals’ aggressiveness. Thus, it is clearly seen that the best evacuation was based on the familiarity of the individual with avoiding aggressive behavior [24].

In 2018, Kontou et al. created a model of crowd evacuation based on cellular automata (CA) parallel computing tool to simulate and evaluate behaviors and distinct characteristics of the people within the area of evacuation, which comprises disable people. A secondary school in the region of Xanthi, which contained disable children, was chosen for the simulation process. With noticing and existing earthquake, the school controlled safety exercise; the overall time of the evacuation was recorded. Finally, the proposed model via the empirical data authenticated and there was a convenience inference to the special area [25]. Table 1 shows information for future use.

The main goals of this paper are: 1) Collecting most of the research works related to various applications applied to pedestrian evacuation using a cellular automata approach in the microscopic classical model. 2) Investigating features, techniques, and implications of the different applications. 3) Learning from 1 and 2 to make a decision to design a new intelligent and reliable model to simulate participants’ appearing emergency behaviors and evacuation efficiency during an emergency evacuation. The model is designed based on an amalgamation of cellular automata, fuzzy logic technique, KNN algorithm, and some statistical equations. Consequently, this paper provides details and specifics in the area of pedestrian evacuation and provides opportunities to the researchers to reach the related information easily and determine their future research directions.

This research study’s contributions are: 1) Utilizing participants’ physical, biological, and emotional elements to apply different speeds for each participant. 2) Incorporating the speed with different elements, such as environment, participants’ distributions, and familiarity of the participant to the exit doors to the evacuation model to improve their appearing emergency behaviors and evacuation efficiency.

The remainder of this paper is organized as follows: Section 2 presents the research method. Section 3 shows the result and analysis of the simulations. Finally, Section 4 concludes the main points and suggests future research studies.

In this research work, the model is based on various scenarios, which have been written and executed through various experiments. Concise of experiments’ results for the scenarios via the developed intelligent model were briefed inside Table 4. For this simulation process, data gathered for 81individuals. As mentioned in the previous sections within the gathering data, individual properties were recorded for each of them. Therefore, these properties were entered into the model, based on these properties individuals were generated and desired speeds were specified for them. These individuals arbitrarily set through the cafeteria. Figure 8 illustrates the participants’ distribution through the first floor.

The above-mentioned scenarios could be briefed in the following points: as mentioned in subsection 4.2, 43.21 percent was male and 56.79 percent was female with (1) One main exit for the first floor, two exit doors for students part, only one exit door to each employee part and staff part, and distribution of the individuals in small and large area of the cafeteria. (2) Two main exit doors for the first floor, only one exit door for each of the student part, employee part and staff part, and distribution of the individuals in a small and large area of the cafeteria. (3) Two main exit doors for the first floor, two exit doors for student part, one exit door for employee part and staff part, and distribution of the individuals in a small or large area of the cafeteria. (4) Three main exit doors, two exit or one exit doors for students’ part, one exit door for each employee part and staff part, and distribution of the individuals through a large area of the cafeteria. Individuals’ familiarity is considered with the points (2), (3) and (4).

8.jpg

Figure 8. Illustrates the participants’ distribution through the cafeteria

The simulation results of the above-mentioned points via our developed intelligent model were briefed in Table 4. Some abbreviations were used, such as C.W.A is a collision with agents, C.W.O is a collision with obstacles, and O.B.D.E is occurring different behaviors during evacuation. Inside the developed our new model, it was appeared that increasing number of room exit doors, main exit doors with familiarity of the evacuees to the exit doors and distribution in a larger area of the floor are the key factors to minimize the congestion, collision and appearing agent’s emergency behaviors that directly have a great influence on improving evacuation efficiency. In the remaining of this section, the result of our model simulation is discussed. For details of simulation results, refer to Supplementary File comprises 36 pages is available in the supplementary Files.

Obviously specifying heterogeneous speed for the participants creates jamming due to colliding agents with each other and with obstructions. This situation was clearly seen in the simulation results. Furthermore, changing numbers of main exit doors and cafeteria’s distinct parts exit doors create variation in evacuation time and appearing emergency behaviors, such as C.W.A, C.W.O, and O.B.D.E. For example, in experimentation of No1C, No2A, they had different numbers of main exit doors, but they had the same distributions, C.W.A, C.W.O, and O.B.D.E were recorded as 27.66 percent, 33.33 percent, and (aside, wait) for No1C respectively, 45.679 percent, 50.617 per-cent, and (aside, wait, help) for No2A respectively. Besides, it was found that with adding a number of main exit doors C.W.A and C.W.O were decreased and different O.B.D.E appeared. Furthermore, agents’ evacuation times were reduced and recorded 0:46:74 for No1C, 0:47:805 for No2A, the time managed as a format of min.sec.ms.

When participants distributed through a larger area of the cafeteria great effect would be created on the occurring behaviors and also on the evacuation time. For instance, in the experimentation of No1A and No1C behaviors, such as C.W.A, C.W.O, OBOE were 61.78, 30.64, and (aside, wait) for No1A respectively, while these behaviors were 27.859, 20.988, and (aside, wait, help) for No1C, respectively.

Because of the larger distribution in No1C collision behaviors significantly decreased between participants and participants with obstacles. Besides that, evacuation time considerably improved and recorded as 1:0:941 for No1A and 0:46:711 for No1C. On the other hand, participant familiarity with the exits doors had a great impact on changing the participants’ evacuation behaviors and evacuation time, for instance, in the experimentations of No1A vs No1B, No3A vs No3B, and No4A vs no4B behaviors and evacuation time, such as, C.W.A and C.W.O changed from 61.728 to 51.852, 65.432 to 30.864, 1:0:941 to 0:47:261 for No1A and No1B, respectively, 23.457 to 16.49, 13.58 to 16.49, 0:41:241 to 0:37:976 for No3A and No3B receptively , and 25.926 to 22.222, 7.407 to 7.407 ,0:40:51 to 0:31:307, for No4A &No4B, respectively.

Furthermore, changing the number of exit doors made a great change in participant’s evacuation behaviors and evacuation time, for example, in the experimentation of No3A vs No1A could be clearly noticed. In No1A which had two main exits and participants distributed through a small portion of the cafeteria, the appeared behaviors, such as, C.W.A, C.W.O, and O.B.D.E significantly more occurred than in No3A which had three main exits and participants distributed through larger area of the cafeteria, the result of this experimentation for behaviors, such as, C.W.A, C.W.O, OBDE, and evacuation time was recorded as 61.728, 30.864, (aside, wait), and 1:0:941 for No1A, receptively, 23.457, 13.58, (aside, wait), and 0:41:241 for No3A, respectively. The result shows appearing behaviors decreased and also evacuation time considerably improved due to an increase in the number of main exits and area of distribution.

From the experimentation result of No1D vs No2D, it appeared that increasing the number of exits is not the only reason to decrease appearing emergency behaviors and improving evacuation efficiency. For example, in the experimentation of No1D, there were two main exit doors, and in the experimentation of No2D, there was only one main exit door. However, the C.W.A is better to be compared to C.W.A in No2D, but C.W.O and evacuation time was better in No2D due to larger distribution.