3. User Behaviour in Terrorist Acts to Model the Evacuation in Outdoor Open Areas

The OA modelling focuses on outdoor spaces and areas surrounded by buildings. Before the attack, users can decide to perform social/leisure/physical activities in the OA and select their position outdoors where the main attractors for (over)crowding (see the scheme in Fig. 3.2) are located. At the same time, these attractors also act as attractors for the attackers to maximize the attack effects being OAs typical soft-targets [11, 33, 47]. First, these attractors are the intended use of outdoor areas [48], by mainly considering pedestrian areas and dehors, open-air terraces of bars and restaurants, open-air market areas, or other (mass)gathering spaces [46]. Second, (over)crowding in the OA can be also due to the intended uses of indoor areas involving possible (over)crowding, such as buildings open to the public and those having a symbolic or cultural value, which also represents an ideal soft target for the attackers [11, 47]. In this case, it could be possible to consider that crowding levels could be reached in front of these indoor intended uses, within their space of relevance (see Chap. 4), e.g., considering that users are waiting to enter [20]. Other buildings which do not host crowding-affected uses (e.g., residential buildings) represent obstacles to user movement by simply bounding the outdoor spaces. Additional components of the OA layout in outdoors to be considered are carriageways and parking lots, in which no initial crowding is considered in view of their use by motor vehicles [46], and obstacles such as monuments, fountains, fences, shrubs and hedges, trees, street furniture, and other fixed obstacles in the OA with protection attributes such as passive and active barriers (engineered planters, wall barriers, low walls, fixed and retractable bollards, heavy objects, water obstacles, jersey barriers) [12]. In this sense, the OA also includes the safe areas, which represent the evacuation targets. They can be defined within the OA (e.g., physically surrounded by fixed protection obstacles), placed in the buildings surrounding the OA (i.e., according to “invacuation” strategies towards protected spaces²), or, at least, represented by the access streets to the OA in view of the need for the users to leave the attack-affected areas [11, 12]. In the CA approach, squared cells with a side of 50 cm can be assumed to represent the OA, being consistent with the general user’s dimensions and ensuring a reliable prediction accuracy [38‐43]. Each of the cells is characterized by a specific typology depending on the aforementioned OA resources (Fig. 3.2). Finally, the OA can be also characterized by the presence of other mitigation strategies, both structural and non-structural, as well as rescuers’ support and evacuation plan (including the coordination of first responders and law enforcement agencies on the site). They can impact the way attackers can provoke damage to the OA and to the users.³ These resources can additionally alter the evacuation path selection by the users and the direct effects on them due to the attack.

The attackers are mainly aimed at maximizing the attack effects on the OA and the hosted users and thus their main task is provoking damage [38], while secondary goals could also concern the rapid escape from the OA after the attack, without being arrested by law enforcement.

The primary issue concerns the initial position of the attackers in the OA, which depends on the intended uses of indoor and outdoor areas in the OA. Considering the OA as an ideal soft target, the main attack attractors can essentially be identified by the most crowded areas and by the areas placed near targets with symbolic value, such as worship, public administration, and cultural buildings [33, 47, 49]. This initial position can be reached before the attack starts or during the attack itself.

Effects and damage, as well as the attackers’ patterns, depend on the specific “modus operandi” and on the number and typology of involved attackers [11, 13, 33].

Bombing attacks could imply “static” effects in the simulation depending on the typology of the bomb and thus on the magnitude of the explosion and the radius of the effects, while direct movement of the attackers could be excluded from the simulation [27]. Similar issues can be linked to Chemical, Nuclear, Radiological, and Nuclear (CBRN) attacks [21, 33, 50], which can also involve a wider urban scenario apart from the OA.

Armed assaults are performed with different weapons [11], and they widely rely on a prey (the users)–predator (the attackers) model, in combination with the “shortest distance strategy”, in which the predator essentially tries to prey on the closest users as the best attack preference [37, 44, 51‐53]. Moreover, the effects on the user and the movement rules of attackers essentially depend on the selected weapon and on its “attack radius”, which is a distance threshold for effective casualties. The overall approach essentially considers the following simulation steps:

Armed assaults with fire guns [11, 51, 54] are characterized by a wide distance threshold. Main behaviours essentially relate to the exploration of the OAs by single or multiple attackers, and their related possibility to remain in effective positions or move along effective paths for a significant time, shooting towards the users. Similarly, attacks with a cold weapon (e.g., knife, sword) can be performed by one or more attackers. The distance threshold for related casualties and the casualty rates depends on the typology of used cold weapons, but general values can range from 0.6 to 1 m radius [20, 42, 44, 51, 55]. This distance threshold can be associated with the probability that the attack can effectively provoke a casualty, in percentage terms [53]. The Terrorism Self-Aid Procedure (TSAP) probability threshold [%] can be hence associated with the users who suffer from the attacker’s action [20], affecting the probability to suffer from the attack as the main resource in the ABM of Fig. 3.1 (see also Sect. 3.4.3). Thus, a casualty is provoked when the user-prey is placed inside the attacker-predator’s threshold and if the user’s TSAP is lower than a considered TSAP threshold. Nevertheless, this TSAP threshold can depend on the weapon typology and on the individual skills of a given user.

An attack with a vehicle running into a target is a typical outdoor attack in the OA [11, 33]. The target can be represented by the crowd (focused on a specific area or dispersed within the OA), a building (i.e., the vehicle moves towards the building façade or entrance), or a specific intended use placed outdoors, especially where (over)crowding levels or symbolic value are relevant. Besides the target, the vehicle driver can essentially adapt the vehicle trajectory to increase damage levels on the crowd. Although a significant lack in current literature is associated with the simulation of such type of attack, the proposed ABM model can be suitable to represent the related dynamics, by simply considering that the attacker corresponds to the vehicle itself and that the movement will be organized according to the possible microscopic trajectory of a vehicle. In this case, the attacker will mainly strike the users placed along the vehicle trajectory, while additional stampede effects could be simulated too [38] (compare with Sect. 3.4.3). The use of a distance threshold and a TSAP can be also considered for the attack with a vehicle. In particular, the distance threshold can be essentially considered equal to about half the vehicle width, thus considering users knocked down by the vehicle.

Attacks by unmanned vehicles/aircraft systems can essentially follow the same general rules of the attacks with a vehicle running into the target [11, 21]. Arson attacks can be essentially modelled according to fire-spreading dynamics [11].

Considering the other main Global Terrorism Database “modus operandi” [11, 56], it could be pointed out that unarmed assaults are specific typologies in which the crowd itself performs the attack as a whole, such as in the case of insurrections. In this case, the users (thus the crowd) and the attackers’ dynamics are essentially overlapped towards a (soft) target. Similar issues are also related to barricade incidents, while facility/infrastructure attacks seem to be out of scope in this model with application to the OAs.

Finally, in risk assessment analysis, the ABM-CA model could also provide a “baseline” scenario condition referred to as the simple evacuation of the OA [52]. In this scenario, the attacker is not directly considered in the ABM and thus no effects of the attack on the users are generated. It hence allows to assess basic interactions between the users and the OA, regardless of the “modus operandi”.

In all the attacks, the modelled emergency scenario implies the evacuation of the OA, since the attack is performed outdoors. Thus, it is considered that “users initially placed indoors do not need to participate in the evacuation process and can simply remain inside the buildings, where they are protected from the accident” [20]. Meanwhile, the initial position of users placed outdoors before the attack depends on the social/leisure/physical activities performed in the OA [46, 48], according to the intended uses of outdoor areas discussed in Sect. 3.4.1. Nevertheless, specific outdoor areas can attract specific typologies of users depending on their behaviours, and individual vulnerabilities and related features (e.g., by age, gender, and motion abilities) depending on their intended use. In this sense, at a broader level, users have to be also modelled to take into account their individual speed depending on age typologies [8, 32, 57], which implies an individual adaptation to the fundamental diagram shown in Eq. 3.1, by considering: (1) the ideal reduction factors on individual speed depending on age reported in Table 3.5; (2) the introduction of an individual random variation in speed (e.g., 0.7 m/s). According to the adoption of grid cells in the OA, the users’ density ρ in Eq. 3.1 can be calculated according to the extended Moore neighbourhood approach [20, 44, 58]. In particular, the approach considers the cells that can be reached by the user i within 1 s of simulation time (as reaction time), at i’s current speed. The density is calculated by excluding cells which are occupied by obstacles to evacuation paths. Moreover, the analysis could be limited to the cells placed within the users’ view cone⁵ and thus along the possible cells placed along the user movement direction, to consider the users’ visual perception domain [41, 59, 60]. The use of this view cone can smooth the individual local trajectories by limiting sudden movements which are not experimentally noticed.

The evacuation start can be performed by users when they are aware of the attack [21]. The signal reception of information by the users led them to perform an initial about whether to evacuate. For instance, input data to this end could be correlated to huge sound levels, presence of smokes, individuation of suspected attackers or injured people, as well as surrounding crowd who has already started running and instructions by first responders. Thus, the start decision depends on the attack typology and related weapons, and it could be differentiated across the OA spaces, also depending on the position of the attack source, especially in case of attacks with cold weapons or vehicles running over the crowd. To consider these phenomena, individual pre-movement time [8, 32] can be modelled, thus including a delay between the attack starting and the evacuation start (e.g., depending on the distance from the attack area, recognition delays of the event). Nevertheless, two opposite but critical conditions could be identified, especially in dense crowd scenarios: (a) synchronous starting of user movement, which increases interactions among moving users; (b) activation of the evacuation start by distance from the attack source, as in the “Mexican waves” phenomenon [61], since moving users can impact those who are still waiting to start evacuating.

Users then start moving to evacuate the OA, by taking into account multiple tasks and resources as shown in Fig. 3.1, while being attracted by a safe area. The elements of reference for these behaviours essentially provoke attractive and repulsive phenomena in users’ local and global paths, which are composed of the selection of different cells describing the OA. Thus, a dynamic floor field model “the willingness to walk” of a user placed in a certain cell towards one of the safe areas, according to a sort of affordance-based approach [40, 43]. Equation 3.2 provides the calculation of the affordance value Aff_c,t [-] associated with a cell c of the grid, at time t, as proposed by previous research [20].

Aff_c,t dynamically changes over the simulation time depending on the composing factors, which are associated with related non-dimensional weights [40] (whose sum is equal to 1).⁶ The probability that a user selects the cell c increases when Aff_c,t increases. According to Eq. 3.2, these factors are:

Aff_c,t varies from 0 to 1, since each composing factor in Eq. 3.2 is based on the normalization rules expressed by Equation 3.3, in which f_c is the value of a factor affecting Aff_c,t considering c, and the subscripts max and min respectively describe maximum and minimum values among all the cells of the OA grid [20].

Figure 3.3 graphically shows the combination between the dynamic and static affordance factors described above, depending on the input scenario at a given time t, by tracing the related maps (the OA is divided into cells) and the overall Aff_c,t map as the overlapping of them (in this case, all the weights are equal to 0.25 to overlap the related behavioural effects. It is worth noting that the factors in Eq. 3.2 could be integrated with attraction effects due to the presence of trained evacuation leaders [51, 65] (thus including attraction rules between pedestrians, rather than just repulsive phenomena as in P_i,c,t). In this sense, affiliative behaviours related to users’ rescuing and support in motion (e.g., users trying to reach other injured users and then moving in close groups) [8] could be simulated according to the same criteria. Moreover, counterforce measures by law enforcement agencies can be also added to the model by considering, for instance, policemen fighting attackers and thus modifying R_c,t and including them as new specific users within the model [15].

A simulation workflow for calculating affordance using the C A approach. The input scenario is a map with overlapping regions representing users and attackers. The affordance factors assessment and final affordance output are represented by C A maps with varying factor values indicated by color intensity.

The surrounding conditions can also lead users to suffer from specific threats. Besides the probability to suffer from the attack (see Sect. 3.4.2), users can stop the evacuation process depending on the probability to suffer from physical contact and to be thus involved in falls [20, 32, 60]. Physical contact can appear evacuation in case of significant crowd density (>3 or 4 persons/m²), of sudden reduction of the motion speed (deceleration > 0.3 g), of users moving in a counterflow, and of individual vulnerabilities (age or motion features related, e.g., elderly and assisted users could be more vulnerable to physical contacts). Probability thresholds to stop the evacuation can be then assigned to each user. In case the threshold is overcome, the user falls to the ground and should spend time rising up and restart moving. Previous works assigned a probability threshold equal to 5% and a random uniform distribution of fall time from 0 to 30 s [20, 32].

Previous works also tried to include “panic” effects within the terrorist act evacuation model [41], but these issues are not considered herein due to the poor validation by experimental-based data. Moreover, fighting behaviours are not modelled in Fig. 3.1 since they are limitedly noticed in real-world scenarios and law enforcement agencies’ recommendations are essentially based on “run and hide” procedures (compare with Sect. 3.2).

From a simulation tool development, the CA model approach defined by Fig. 3.2 takes advantage of simulation time discretization [20, 32, 43]. The time step between two consecutive time t and t + 1 can be modelled depending on the maximum user speed, so as to represent the quickest evacuation process within the simulated agents [20]. Asynchronous update rules for user movement can be then considered, assuming: (1) a random selection in the users’ simulation order at each step; (2) that each user can wait or move one cell per step by selecting the next one within the neighbouring ones placed along the movement direction and inside the view cone.

Finally, when reaching a safe area, the user exits from the simulation. Otherwise, the users can be removed from the model in case they suffer from the attack (being affected by casualties depending on TSAP, compare with Sect. 3.4.2) or when the maximum simulation time is reached.