A Wireless Bracelet: Self-Organization of Pedestrians for High Density Prevention in Crowds

(1)

Faculteit der Natuurwetenschappen, Wiskunde en Informatica Graduate School of Informatics, MSc Computational Science Universiteit van Amsterdam Department of High Performance Computing In Multidisciplinary Research Faculty of Informational Technologies and Programming Saint Petersburg

National Research University ITMO

A Wireless Bracelet: Self-Organization of Pedestrians for High Density

Prevention in Crowds

Author:

Jiaqi Liao

Supervisor:

Michael Lees

Sergey Kovalchuk

August 13, 2015

(2)

Abstract

As the frequency of crowd stampedes increases in this century, there appears an urgent de-mand for an efficient tool for crowd-disaster prevention. Although there exists much research effort on this field, such as crowd management and investigation of crowd hazards and crowd dynamic, there is no an adequate crowd-managed strategy.

Most of the proposed solutions for human behavior detection are based on complex cal-culations of video or motion graphics. Some methods are functioning by regions using other sensors, and the crowd is managed by event managers. There is little research which takes a self-organized possibility into account.

In this thesis, we develop a wireless bracelet for spatiotemporal estimation of crowd density within range. The real-time density assessment and message propagation through the ”device network” improves information accessibility and situation awareness in pedestrian crowds, to achieve an early warning function. We compare the effectiveness and power consumption of the bracelet with different configurations, such as communication range and transmission condition.

(3)

Acknowledgements

First of all, I would like to express my deepest appreciation for the opportunity of studying in the Dual-Degree Master Programme ”Computational Science” between ITMO and UvA, provided by Alexander Boukhanovsky, the director of Department of HPC at ITMO University. It allows me to experience different but both fantastic life in Russia and Netherlands. This amazing experience raises my enthusiasm for multidisciplinary research, especially in urban computing and data anal-ysis.

Secondly, I would like to thank my teachers at ITMO, especially Alexander Boukhanovsky and Alexey Dukhanov, who always try their best to support me during my study and Sergey Kovalchuk for his patient guidance on my research in the first year. Also, I am grateful to my friendly group-mates, who have been generously helping me and delighted my life in Russia.

Thirdly, I would like to present my intense gratitude to Michael Lees, my supervisor at UvA, for his inspiring advices on my research. He provides me the useful simulation platform with sourcecode, which allows me to focus on the analysis. Without him, I would not have found this interesting topic and completed this thesis.

Lastly, special thanks and love to my family for their absolute faith and support, emotionally and financially, on my decision to study further abroad.

(4)

Chapter

1

Introduction

Since 2000, the global population has increased with a growth rate of under 1.3% per year [1].Till 2015, world population has reached 7 billion and is still increasing. Meanwhile, urban population in 2014 accounted for up to 54% of the total world population, and continues growing [2]. It brings out the higher population density of urban area and raises a big challenge on crowd safety. With the raising urban population density, there are more chances for people to gather together, with an accordingly higher probability of crowd disasters, especially in demonstrations and religious ceremonies.

Until now, about 30 human stampedes in the 20th century and 33 human stampedes in 21st cen-tury have been recorded [3]. One of the grievous crowd disasters last year is the 2015 Shanghai New Years Eve stampede, in which at least 36 people were killed. It was caused by neglected overcrowding and triggered by a collision on the stairway, about which people who were moving toward the stairway were not informed and continued pushing forward [4]. Another well-analyzed crowd disaster was the annual ceremony Love Parade held in 2010 in Germany. 21 people died in this event as a result of high crowd pressure. Due to the lack of a sufficient solution for crowd management, this event has been permanently cancelled since 2010 [5]. Stampedes in religious festivals often cause even more severe consequences, however, these events cannot be cancelled. In 2013, two human stampedes, Kumbh Mela stampede and Madhya Pradesh stampede, killed 36 and 115 people respectively [6, 7].

With regard to the cause of these disasters, it is usually the inaccurate estimation of facility ca-pacity and inadequate event instruction. In practice, a lot of approaches are adopted for crowd disaster prevention, such as using a guideline or prior event simulation. Although there exists much research effort on crowd analysis and disaster prevention, most of it focusses on the pre-vention strategies in advance, such as facilities design, guidelines formulation, forecast of crowd behaviors. Normally these methods fail to consider all the unexpected cases, due to the complex-ity of human behavior. There is little research on real-time crowd management, especially in a self-organization of pedestrians.

This thesis proposes a model for crowd management, which is a wireless LED bracelet for real-time crowd management. The low-energy and low-memory bracelet uses the device introduced in [8] as prototype, which can observe and communicate with its surrounding devices within a certain range, and inform the wearer about crowd density by LED light. The majority of peaceful-gathering disasters happen because of inaccessible information. I.e., entering pedestrians do not witness and notice the overcrowding ahead and continue pushing forward, increasing the size and pressure of the congestion. The proposed bracelet works on high-density detection and propagation, so as to warn its wearer off the unnoticed danger. In this thesis, we aim to investigate an optimal configuration of this device, such as the communication range, the propagation range, the power consumption. The remainder of this paper is structured as follows: Section 2 starts with an overview of the com-parison of several crowd behavior models, then introduces some current results for crowd control,

(6)

and summarizes the limitations and challenges in this area. Section 3 provides a concrete insight of Social Force Model [9], bracelets model, and the power consumption model we used. Section 4 discusses the experiment design and presents the result for three scenarios. Section 5 concludes with our results and future work.

(7)

Chapter

2

Literature Study

This section starts with the introduction of some pedestrian and crowd models. Then it describes some common features and causes of crowd stampedes, using Love Parade Stampede as a case study. Next, it presents current work for abnormal crowd behavior detections and disaster preven-tion and their limitapreven-tions, including a model, to which is referred by this thesis .

2.1 Pedestrian and crowd modeling

Pedestrian dynamics is an area that fascinates scientists and researchers in many domains, such as sociologists, psychologist, and computer scientists. The goal of pedestrian crowd study is mainly to develop a level-of-service concept, mass event planning, design of pedestrian facilities. Traditional methods for pedestrian analysis are mostly based on direct observations, such as photographs and videos, however, the results are usually dependent on specific conditions, such as facilities structure or buildings architecture.[10] As the development of computation technologies advances, a lot of general and quantitative crowd models have emerged.

In [11], the author summarized and categorized several existing crowd models and modeling tech-niques according to the size of crowds and the time-scale of the interested phenomena. Huge-size crowd modelings usually consider the horde as a whole and focus on the global trend of the crowd motion, while for small or medium size of crowd, emphasis are put on individual behavior and interactions. In this thesis, the developed bracelet works for individual according to spatiotempo-ral density, e.g., the tempospatiotempo-ral number of neighbors for each pedestrian. Hence, we centre on the medium-size crowd simulation models, among which agent-based simulation is the most natural way to model pedestrians.

In an agent-based model, each agent has its own characteristics, such as mass, weight, walking speed, psychological status, cognitive ability. Coordinated with self-organized effect, individual reacts to the local environment independently. With extra domain knowledge of social crowds, we can easily implement such simulation models, so as to concentrate on analysis of internal factors and coordination problems of hordes. In other words, the model is flexible to extend with addi-tional features [10]. Therefore, it become one of the most popular approaches for crowd simulation on micro-investigation of pedestrian behavior and interactions. There are various self-organized pedestrian models[9, 12], out of which Helbing’s ”social force model” may be the most wide-spread one.

Helbing’s social force model [9] is applied in our crowd simulation. It is the most popular model for investigating pedestrian dynamic in recent years, because of its relatively accurate emulation of normal pedestrian behavior [13]. In this model, each individual is considered as a particle, which adapts its velocity to its desired speed, influenced by a so-called ”social force” from other particles

(8)

or obstacles. Here ”force” is individual’s internal motivation from a psychological perspective, i.e attractive forces to the destination and repulsive forces from its neighbors. There are three essential features observed to determine the next step movement of a particle [9]: (i) the preferred velocity of motion with an unaffiliated fluctuation due to avoidance process, (ii) a minimum distance that a pedestrian keeps to others to retain a free space around him, (iii) the destination, such as concert stages or exits. As a result of these observations, it can be naturally simulated by a self-organized method. The detailed model would be described in Section 3.1.

Another popular pedestrian-animation model is Hierarchical Model proposed by Musse S.R. and Thalmann D [14]. This model analyzes pedestrians under 3 layers: individual, group, and crowd. What’s more, it presents pedestrians reaction with various levels of observations, including intem-perance, rules announcement, and external-guided behavior. This model is also simulated by a agent-based way, while it is more complex than the Social Force Model. It indicates some interest-ing factors in groups and crowds, which in practice is more realistic. Our bracelet is worn by each individual, and works depending on only the density of each single pedestrian. This simplicity allows us not to put emphasis on the individual reaction and the pedestrian behavior pattern. Social Force Model provides adequate performance for our experiments, so we do not choose the complex Hierarchical Model.

Recently, psychological factors, such as emotions, have drawn researchers’ attention. Models con-cerning about emotional influence are usually operated in some urgent situations, such as congestion and collision, escaping panic, a long-time waiting queue. These kind of models are not applicable for our assumption.

2.2 Analysis of crowd disasters

[15] investigated the important role of pedestrian heuristics in crowd disasters. It concluded that the physical interactions between individuals may occur unintentionally in the case of overcrowding. The combination of pedestrian heuristics with body collision often generates crowds disturbance and disorder, which often finally trigger stampedes. As Love Parade Stampede is a representative crowd hazard, which happened without obvious hint out of peaceful pedestrians, [16] analyzed the cause of this disasters. It provided several features of crowds that might lead to stampedes:

1. A sort of ”queue effect” often leads to denser and denser queue of people over time [16]. It is a common phenomenon that a lot of new-coming pedestrians push into the crowd, inducing heavier pressure inside, rendering body contact among people. It often furthermore triggers collisions when a queue of people lose their patience because of long-time delay or the threaten stress [17, 18].

2. Under the condition of dense stress, people still behaves politely and rationally, however, with our an indication of progress for delay, they may lose patience and some of them will ignore the regulations and start to find a way out. For instance, those standing around the staircases, slopes, walls, intend to evacuate themselves by climbing walls or exploring alternative ways out. This result in the phenomenon that a mass of population push forward these ’exits’ [16].

3. At a extremely high density crowd, people under high pressure are suffered from accumulated fluctuating forces by nearby individuals. By these random forces, they can hardly keep their balance and determine their movement freely, in other words, their direction of motion is dominated by the surroundings. This often implies the beginning of ”crowd turbulence”. 4. Under a ’turbulence’ state, people out of balance will easily stumble and fall down, and others

are forced to step on them in order to keep the relative balance of the crowd. This is how a tragedy starts [16].

(9)

The author also discuss some early warning sign of this disaster, which should catch people’s at-tention, and further actions should be taken to forbid worse consequences [16]. For example, when a large number of pedestrians gather and accumulate in certain regions, and the density of these regions keeps increasing, limiting entering flow should be immediately taken into action to ensure the population will not exceed the capacity. Similarly, when jams of pedestrians appear and the size of jams keeps growing, which usually occurs near the exit, communication with the crowd and information accessibility play an important role for efficient and safe evacuation. Other obvious hints, such as people falling, people shouting, or disrespecting the rule, should draw attention of the event managers.

Despite this prove to be a detailed and valuable study of the disaster, Love Parade is stopped holding due to the inability of crowd control. It reveals that there is no sufficient solution for such an event yet.

2.3 Existing model for crowd monitor

In order to prevent pedestrian hazards, it is crucial to understand and recognize crowd abnormal phenomena. Many researches on abnormal behavior detections have been proposed these years. Pedestrian density [19, 20, 21] and anomalous change of velocity are two ordinary features to rec-ognize abnormal situations. Some other observations, such as change of optical flow [22, 23], Radio Signal Strength [24, 25, 26] are used to interpret the aforementioned features.

Two ways are mostly adopted to extract these features: video [27, 28, 22, 23] and other sensors [29, 24, 30, 25]. Many researches work on abstracting main factors from video to investigate crowd motion, because video is the most common way to trace crowd behavior and it contains rich precise information. [22] captured the anomalies by the likelihood of optical flow extracted from video judged by a HMM. Instead, [23] calculated the force flow using social force model and probes abnormal phenomena from video by a bag of words method. However, computing the video im-age and analyzing the pedestrian state on real time is still challenging due to the computational complexity and the insufficient coverage of cameras. Furthermore, it can hardly run in a low-light condition. And it is not flexible to install cameras everywhere.

Due to these limitation, other sensors are chosen to collect needed information. Wireless sensor [24, 30, 24, 26] is recently popular in collecting information to monitor pedestrian crowd. [24, 24, 26] use fixed-located radio signal strength to demonstrate the crowd density of an area. This can work in dark condition automatically. However, for some activities hold in wide places, high density of crowd can hardly be managed and controlled by regions.

Therefore, there emerges some geographically-independent solutions for crowd observation and man-agement. For instance, [31] developed a method for crowd behaviors recognition by mobile sensor coordinated with pattern analysis. In their work, they used a simple local pairwise calculation of signal combined with a global graphic clustering to investigate some common crowd patterns, such as queueing, clogging, group formation, which can be further studied for improving situation awareness, event direction or event evacuation under urgent conditions. It contains a process of learning-by-demonstration which is mapping the local signal with global graphic extracted. How-ever, the abstraction of graphical factors and the learning process are too time-consuming while a real-time monitor is needed.

[32] introduced another sensor-based density-detection model, in which an mobile-embedded sen-sor continuously collect nearby pedestrians’ information and periodically upload it to closest cen-tre(base) device. The author focused on the collective cooperation among devices and the efficiency of data transmission instead of density control. After all, It gave abundant inspiration for crowd information collection. [33] proposed a sensor network for evacuation out of a building, which puts centre on path-planing according to density of certain spots such as a exit or corridor as well as

(10)

their distance to their destinations. Each node can communicate with nearby devices for exchang-ing information about the contiguous crowded status and calculate the distance to the exits, based on which it then guild the wearer to leave with safety. This model emphasizes on the evacuation strategy, however, a map of the event zone is pre-required.

[30] presented a vibro-tactile LifeBelt, with 8 embedded vibrator elements, as a guidance for emer-gent panic pedestrians evacuation. Separated sensors embedded into the fabric belt are connected to a master controller, from which each element receives and follows orders to instruct the wearer’s movement. It also illuminated the possibility of a wearable wireless sensor for crowd regulation. We seek to find a way that peaceful-crowd density can be controlled by a simple but efficient strategy to prevent crowd disorder and turbulence. Despite research on pedestrian behavior plays a vital role in crowd analysis, it is to some extent regarded dependent on pedestrians’ experience, habit and culture, and its psychological states [16]. It needs a enormous dataset to explore in this field. Alternatively, there exists solutions for estimation of crowd without geographical or cultural restriction.

Here we would like to introduces a low-cost wearable wireless devices developed in [8], that used to keep from high density by continuous pedestrian counting and situation propagation. As soon as the number of ambient devices is greater than a given threshold, it is considered in a dangerous condition. In the meantime, a ’danger signal’ would be broadcasted through this network within region, so that people would be informed when they are close to a high dense region. This model allows wearers to direct themselves wandering by an organized way, according to the instruction of the devices. It is a new self-organized model, compared to the case without this device, the wearer have more information than alone. Our study is inspired by this model. The detailed model description will be given in next section.

(11)

Chapter

3

Model Description

3.1 Social Force Model

Social force model [9] is formulated by Helbing in 1995, and since then it has become the primary applied collective pedestrian model thanks to its well-extendability and lifelike-pedestrian anima-tion. In this model, each individual is considered as an intelligent agent, which makes decision of next step according to its observations and motivation in a self-organized way. The intention of movement forms and is influenced by a so-called ’social force’, which contains an attractive force to its destination and repulsive forces to keep a minimum distance from other agents and obstacles. In overcrowding or emergent cases, such as fire evacuation, physical forces occur and should be taken into account. In practice, attractive effects among members of a group exists and can be added to the fundamental model. Therefore, in the simulation, each agent takes actions based on some relatively simple calculation, which mimics pedestrians’ immediate reaction to the environment ’automatically’ instead of taking complicated decisions. The formulations of the basic social force model that determined the pedestrians’ motion are introduced below. Note that all the formulations and equations are given in [9, 16].

First of all, each pedestrian tends to move toward its destination in a straightforward way, namely, the shortest possible way. Considering this way consists of several edges ~r1

α, . . . , ~rαn:= ~rα0, because

of avoidance effect. If the next edge to reach is ~rk

α, the agent’s desired direction ~eα(t) of motion

will be ~ eα(t) := ~ rαk− ~rα(t) k~rk α− ~rα(t)k (3.1) Where ~rα(t) denotes the temporal position of the agent α at time t, and ~rαk refers to the goal,

toward which the agent moves.

Without any interruption by other pedestrians or obstacles, the pedestrian is assumed to walk in the direction ~eα(t) with a preferred speed v0α. Influenced by the individual fluctuation and the

repulsive force from other pedestrians or obstacles, the actual velocity ~vα varies with respect to

the desired velocity ~v0

α := v0α~eα(t). The pedestrian α will adapt his actual velocity ~vαto its desired

one ~v0

αwithin time τα. Given a homogeneous population, the additional force on pedestrian α can

be described by a acceleration term. ~

F_α0(~vα, v0α~eα) :=

1 τα

(v_α0~eα− ~vα) (3.2)

In this model, the interaction force between pedestrian α and pedestrian β depends on the distance kdαβk between them and their velocities, which furthermore determine the pedestrians motion. The

repulsive force from pedestrian β is represented as gradient of a repulsive potential Vαβwith respect

to vector ~dαβ pointing from β to α. Vαβ is a exponentially decaying potential, which formulates

(12)

circular one allows us to take the velocity of pedestrians into account, which is more realistic. So, the assumed formula is given below.

Vαβ(bαβ) = ABe−bαβ/B (3.3)

Here A refers to the interaction strength and B reflects the interaction range. bαβ denotes the

semi-minor axis of the ellipse, and it is specified by 2bαβ=

q

(k ~dαβk + k ~dαβ− (~vβ− ~vα)∆tk)2− k(~vβ− ~vα)∆tk2 (3.4)

where dαβcorresponds to their distance.

The repulsive force is associated with the above feature via fαβ( ~dαβ) = −∇_d~_αβVαβ(bαβ) = −

dVαβ(bαβ)

dbαβ

∇dαβbαβ( ~dαβ) (3.5)

where ∇_d~_αβ refers to the gradient of Vαβ with respect to ~dαβ. According to the chain rule, we can

have the following precise formula fαβ( ~dαβ) = ABe−bαβ/B· kdαβk + kdαβ− yαβk 2bαβ ·1 2( dαβ kdαβk + dαβ−yαβ kdαβ−yαβk ) (3.6)

where yαβ= (vβ− vα)∆t. In our simulation, we take ∆t = 0.05s.

3.2 Device Model

Based on the case study of Love Parade [16], physical force between pedestrians elicited owing to the overcrowding leads to disorder and jumble, which normally result in stampedes. It is crucial to elaborate a strategy to control the mass density.

A wireless bracelet with an affiliated sensor provides a method of real-time crowd management based on mechanical estimation of pedestrian density. It takes Angela’s device [8] as prototype, which is designed to be a low-power and low-memory wearable wireless device. The bracelet sim-ply counts the number of neighbors within a specified range r, and concludes the safety condition around it with this measurement. After that the attached LED light changes its color, i.e red or green, according to the density to steer the wearer’s movement, e.g stop walking. The interactions among bracelets, meanwhile, composes a propagation network, through which the ’high density’ can be relayed further to relevant individuals.

Here we clarify some parameters definitions and specifications:

• Communication range r: A bracelet senses and communicates with devices within radius r. In the model, r is valid from 1m to 8m.

• Increase range threshold Ninc: Within valid range, if the number of detected devices is

smaller than Ninc, r will increase by 1m.

• Decrease range threshold Ndec: Within valid range, if the number of detected devices is

greater than Ndec, r will decrease by 1m. (See Figure 3.1)

• Neighbors threshold Nr: For each communication range r, there is a corresponding neighbors

threshold Nr. The configuration of these Nr is specified in Table 3.1. When the number of

neighbors is greater than this value, it is regarded in dense region, where the communication range r will decrease to minimum, i.e., 1m in our case. Note that the neighbors threshold is not proportional to communication area. Otherwise, when the communication area is wide, an uneven-distributed pedestrian hordes may cause the underestimation of the density in some area. Inadequate assessment sometimes occurs due to particular structures of places, especially under a tunnel scenario. (See Figure 3.2)

(13)

• Danger message threshold Nmsg: If the number of messages a device receives exceed this

threshold, the device is suggested to be in unsafe condition.

• Maximum hops Hmax: A source massage can be transmitted within no more than this number

of hops through the device network.

To sum up, the working process of the bracelet can be divided into three stages: Neighbors detection, Density propagation, Information display.

Neighbors detection Assuming the power is stable and reliable, the device senses its surrounding devices within communication range r at a specific time interval tp (In our experiment,

tp = 0.05s). If the number of detected devices exceeds a predetermined threshold Nr, it is

considered in a unsafe state. To achieve an early warning effect, i.e., the message should be transmitted immediately and far. This requires the giant component cover a sufficient fraction of network. In the case that r is extremely small, a bracelet can observe merely those who are standing very close to it, where apparently the giant component will be small. On the other hand, if r = ∞, every individual is connected to others, forming a fully-connected network, so that the fraction of devices belonging to the giant component is one. The greater the r is, the greater the giant component become in the network. As long as it is in a compact space, however, only nearest neighbors(in the boundary of the component) are contributed to build a bridge between two components, while additional communication with further nodes consumes unnecessary energy. Therefore, instead of a fixed range r in [8], a variable one r(t), which changes according to temporary number of neighbors, is applied in our model.

Communication range r 1 2 3 4 5 6 7 8

Density threshold Nr 7 28 63 112 175 252 343 448

Table 3.1: Device neighbors threshold Nrfor each communication range r

1 x y1 y2 1 6 7 2 6 28 3 6 63 4 6 112 5 6 175 6 6 252 7 6 343 8 6 448 N u mb e r o f n e ig h b o rs N Communication range r 1 2 3 4 5 6 7 8 7 28 63 112 175 252 343 448

Increase and Decrease range threshold [Ninc, Ndec] Neighbors threshold Nr

When N >= Nr, r decreases to the minimum range.

When N >= Ndec, r decreases by 1.

When N <= Ninc, r increases by 1.

s

Figure 3.1: Communication range changes according to communication threshold and device threshold.

Density propagation A device merely knowing its own density does not give much more infor-mation than a human observes alone [8]. To improve inforinfor-mation accessibility in terms of unwitnessed danger, a feature of message propagation plays a significant role in this bracelet. As long as a device is suggested unsafe, the identity of this device will be propagated to relevant devices. In [8], a message can be relayed to its neighbors and further within 2 hops.(See Figure 3.3) A predetermined maximum hops for transmission allows the device

(14)

Figure 3.2: Communication range exceeds the width of the tunnel.

aware of the distance to the danger spot, however, on the other hand, it limits further and sufficient information awareness if necessary. Considering the case that congestion occurs at the centre of a relatively large crowd, only those near to this high-density area will receive a warning message because of the limited transmission range. At most time, people in the boundary of mass will not obtain the message and therefore relay nothing to the new-coming pedestrians. Knowing that the entering-flows have higher autonomy, namely, a broader free space and directions to head to than the ones entrapped in the mass, without a early warning signal, these pedestrians will continue crowding into the density, accordingly, enlarge the size of crowd. It is reasonable to believe that greater size of the mass renders greater challenge for crowd management. Hence, it is vital to come up with a strategy to transfer a danger signal out of the mass. And the process of propagation is supposed to adapt to its temporary density without taken the space structure, e.g., length of the tunnel, into account. In our model, besides primary source which indicates high density around it, a secondary source is produced and propagated when the number of received message is higher than Nmsg.

Information display As long as the bracelet is regarded in unsafe status as a primary or sec-ondary source, it will turn its color to red. The wearer ought to stop moving. Otherwise, the wearer can continue walking with green light.

3.3 Power Consumption Model

In our experiments, power consumption is calculated based on a classical model [34], in which there are three main kinds of energy consumed during the transmission process: energy to receive the data, energy to amplify the data signals and energy to transmit the data.

The following formulas are provided by [34].

ET X(K,d)= ET X−elec× K + εamp× K × d2 (3.7)

ERX(K)= ERX−elec (3.8)

Where:

ET X(K,d) : Energy consumption of sender for K bits for distance d

(15)

1 1 2 3 3 4 4 r

Figure 3.3: Message is transmitted with maximum 2 hops [8].

ERX elec : Energy consumption of sender for message transmission of 1 bit

ET X elec : Energy consumption of receiver for message transmission of 1 bit

εamp : Energy consumption by amplifier

K : Data size in bits

D : Distance between two nodes

Here εamp = 100pJ/bit/m2, ERX elec = 50nJ/bit, ET X elec = 50nJ/bit [35]. For comparison

purpose, we simply set K = 1, i.e., the message for each transmitted message is 1 bit. For example, when a message indicating the density is transmitted to its neighbors by one hop, the sender consumes

Esend= (1bit × 0.1nj/bit/m2× d2+ 1bit × 50nj/bit) × N (3.9)

Where d denotes the distance between two devices, and N refers to the number of neighbors receiving this signal. For each neighbors, the following about of energy is consumed.

Ereceive= 1bit × 50nj/bit (3.10)

Although the concrete amount of power consumed by this process depends on the hardware con-figuration of real devices, the data size, the communication methods, etc., we can still estimate the optimal configuration of the devices with respect to power consumption by comparison using this simple model.

(16)

Chapter

4

Experiments

In this section, simulations are applied to investigate the effectiveness of the aforementioned devices. Due to lack of real data on crowd behavior and the difficulty of data collection and extraction, we utilize simulation of social force model to generate experiment data.

4.1 Experiment Tools

All the simulation is running on MASON [36] simulation toolkit. It is a fast discrete-event mul-tiagents simulation library written in Java, which provides flexible configuration to control the simulation process, such as simulation time interval, maximum number of time steps, and so on.

4.2 Experiment Design

As described in the previous section, the object of the proposed bracelet is to prevent high den-sity in crowd by denden-sity detection and awareness. Learnt from the pass crowd hazards, an early warning strategy plays an important role in crowd disaster avoidance. Hence, we concentrate on the realization of early warning for this device instead of the density area overall.

In the experiments, we compare and examine effectiveness of the bracelet model under differ-ent device configurations, including communication range threshold [Ninc, Ndec], danger message

threshold Nmsg, maximum hops Hmaxby the density distribution of the interested area and power

consumption of the bracelets. We assume a optimal characterization of parameters exists for all cases.

The assessment of the bracelet is operated as following. Firstly, to allow the message to be spread further enough, it requires a minimum degree of network connectivity(i.e., the size of giant compo-nent). Such condition is associated with the communication radius r of the devices. As r changes depending on the relation between its current neighbors count and range threshold [Ninc, Ndec], we

compare the number of dense area and power consumed given different pairs of thresholds. Sec-ondly, we estimate the effectiveness of a function of a secondary source transmission, and emphasis is put on the range of information transmission, given different danger message threshold Nmsg

as well as maximum hops Hmax. Experimental assumptions and estimation approaches are given

below.

Communication range Given different pairs of initiation of Ninc and Ndec (Table 4.1), we

presume that a bigger pair of thresholds lead to a further average sensing range, and a corre-sponding bigger giant component. Nevertheless, the safety and overall density are supposed not to be significantly affected. Here we set Nmsg=3 for all scenarios.

(17)

Danger message threshold With a minimum fraction of devices covering the giant component, the message can be relayed within range. In our model, two sorts of message sources are generated and propagated: primary source and secondary source. We focus on the effect of the secondary source, which is associated with Nmsg, and performance is investigated by

comparison of density distribution and power consumption under various Nmsg.

Maximum hops Each unique density message, including primary and secondary source, can be relayed within a fixed maximum hops Hmax. In principle, a greater Hmax allows the

message to be propagated further and, therefore, results in a safer environment. However, the power are believed to grow accordingly, and an optimal initiation of Hmax is dependent

on the area structure. Therefore, we do not investigate this feature independently, instead, it is mentioned for comparison and coordinated with the message threshold Nmsg to achieve

an optimal performance.

Experiments are applied on 3 scenarios: a square exit scenario, a tunnel scenario with high gener-ating rate of pedestrians, a tunnel scenario with low genergener-ating rate of pedestrians.

Exit Scenario Exit/Entrance is the spot where often gather pedestrians easily and quickly in most event area, especially with single-exit yard. Therefore, it is one of the scenarios that are often chosen to investigate in the domain of service of level or crowd disaster prevention. Here we apply the experiments on a scenario of a single-exit square, where there exists 200 agents at initial time and continues entering 4 agents per second (See Figure 4.1).

Tunnel Scenario Several crowd disasters have occurred under tunnel situations [37, 16], because entrance/exit is often set to be a tunnel structure. Some did not happen in a tunnel, but similarly within a spontaneous-formed long queue of pedestrians [4, 38], in which the entering flows were not aware of the excessive congestion ahead. When the queue is so long that cameras cannot trace, an early warning system is required. In this scenario, it aims to warn the entering flow of a high dense crowd ahead efficiently, so that to avoid an over-crowded situation. Here we apply an enclosed tunnel scenario for the our examinations on the efficiency of information propagation, since it is able to quickly form a congestion at the end of the tunnel. Then, pedestrians are generated following a prespecified frequency with slight variation (See Figure 4.2). In the following experiments, we apply two tunnel scenarios with continuously generated pedestrians with high-frequency (3.5 agents/second) and low-frequency (1.5 agents/second).

4.3 Results

4.3.1 Variable Communication Range

Firstly, we do experiments on a tunnel scenario with agents which are periodically generated with frequency about 1.5/s (1.5 people per second). The result is for 3000 time steps, i.e., 150 seconds. Regarding the estimation of density distribution for all the following experiments, we consider a 1m2 cell unsafe if there are more than 4 pedestrians in this area [16].

We presume only the giant component instead of safety is significantly influenced by the commu-nication range, determined by Ninc and Ndec(See Table 4.1). Figure 4.4(left) shows the number

of high density cells in the experimented area over time, given fixed r and variable r under three aforementioned scenarios. In all three cases, as expected, there is no meaningful influence on the density given threshold up to [7,10]. From [8,10] to [9,12], the density is slightly improved, while still increasing over time.

Figure 4.4(right) exhibits the trend of average power consumption and giant component fraction under various configurations of Nincand Ndec. Note that the agents have been generated through

the simulation time, i.e., they have various length of life, therefore the average of them does not tell the approximate value of consumed energy for a certain length of life time of a device. With

(18)

Figure 4.1: An Single-Exit Square.

regard to the average giant component, the initial giant component coverage depends on the initial density. For example, in the exit scenario, there are 200 pedestrians in the observed area at the beginning, therefore, the initial fraction of devices within the giant component is about 70%.(See Figure 4.4c(right)) To estimate the effectiveness of the determined configuration, we focus on the improvement it brings. As shown in Figure 4.4a(right), given Ninc=3 and Ndec=6 in the tunnel

scenario with low-frequency entering pedestrians, the power consumption is slightly lower than the one given fixed communication range, while it gives much greater connectivity of the network. It is concluded that fewer danger messages are propagated while the density stays almost the same (See Figure 4.4a). A network connectivity of higher than 95% can be reached given from Ninc=5

and Ndec=8 without radically increasing energy consumption compared with applying [4,7].

Af-ter that, the power consumption makes visible growth from [6,9], and increases significantly from [8,11]. Detailed giant component over time for each configuration in this scenario is shown in Fig-ure A.1. With fixed communication range, the giant component is increasing, while it can reach to the coverage fraction of about 0.7 from 2000 to 3000 time steps. Given Ninc=3 and Ndec=6, the

giant component fluctuates. At the beginning of simulation, the giant component is at the back side of the tunnel because of the increased communication range (See Figure 4.3a), and then when a number of agents accumulate at the end of tunnel, the giant component will always include these agents (See Figure 4.3b). A relatively stable giant component fraction can be achieved from [5,8]. In terms of the power consumption for each device, Figure A.2 presents the consumed power at 3000th time step.

Ninc 3 4 5 6 7 8 9 10 11

Ndec 6 7 8 9 10 11 12 13 14

Table 4.1: Communication range thresholds

Secondly, we applied the same experiment with a bigger agent-generated rate, which is 3.5 people per second. The experiments are run for 1500 time steps, which is 75 seconds. Figure 4.4b(right)

(19)

Figure 4.2: An Enclosed Tunnel.

shows that the a similar density trend for this scenario. The results show that the configuration of Ninc=4, Ndec=7 gives a relatively stable giant component. It is reasonable that the denser

crowd allows the communication range to adapt to the density better than a loose network. The corresponding power consumption reflects an almost the same result as for the previous experiment (See Figure A.4 for details). Threshold [5,8] does not increase the power consumption significantly. Therefore, we conclude that thresholds [4,7] and [5,8] optimal options for this scenario.

Thirdly, a single exit square is applied. We place 200 pedestrians inside the ground at the initial time then continue generate new pedestrians to allow the congestion to be formed quickly. The results of density and power consumption can be found in Figure 4.4c. Due to the exiting of pedes-trians, the density does not increase with a high rate as closed-tunnel scenario. The trend of density under different Ninc and Ndec is similar to the tunnel scenario. The difference is that with fixed

sensing range, the average proportion of giant component can reach approximately 70% because of the exiting 150 pedestrians in this area heading to the exit. Detailed giant component proportion over time and power consumption at last time step for each device can be found in Figure A.5 and Figure A.6 respectively. In about 1 minute from the beginning, the proportion slightly increases as more pedestrians enters and herds into the crowd. And we can observe that the average fraction of devices covering the giant component can reach above 90% given [3,6] and above 95% given [4,7]. In this scenario, given threshold [3,6] or [4,7] will not consume more power than fixed sensing range, this is possibly due to the slightly improved density (See Figure A.6). In this scenario, [4,7] and [5,8] are the best options with regard to network connectivity and energy consumption respectively. Based on the above experiments, we conclude that the network connectivity can be significantly improved by the various sensing range compared to the fixed one. With higher density of crowd, the proportion of giant component within the network is greater than the crowd with lower den-sity. Thus, to find an optimal configuration of sensing range threshold, entering and exiting rate of pedestrians should be taken into account. From the above results, we can also obtain that [4,7] and [5,8] render ideal performance for these 3 scenarios. Thus, we apply [5,8] to the following experiments.

(20)

(a) At the beginning, the giant component (in blue) is formed from the back of the tunnel.

(b) A stable and normal state of giant component at the end of the tunnel.

(21)

(a) Tunnel scenario with low-frequency generating agents

(b) Tunnel scenario with high-frequency generating agents

(c) Exit Scenario

Figure 4.4: Density and consumed power given various communication range thesholds Ninc and

Ndec. Left: number of dense cells over time given different Ninc and Ndec. Right: Average power

(22)

4.3.2 Message Propagation Threshold

We investigate the effectiveness of the bracelet with various massage propagation thresholds Nmsg.

If the number of message the device receives is more than this value, the device itself will become a message source to propagate its own identity through the network. We tend to apply this strategy to a almost fully-connected network, so we choose Ninc=5 and Ndec=8 according to the previous

experiments.

Since the giant component should not be associated with the message propagation threshold, we only present the density distribution and the power consumption. Here we have two assumptions, (i) The higher maximum hops is chosen, the higher power will be consumed, however, the lower density will be achieved; (ii) Greater message threshold, namely the stricter requirements for the device itself become a new source, will cause lower power consumptions for propagation, never-theless, the higher density of crowds. In both assumption, there exists a tradeoff between power consumption and density. Furthermore, the power consumption is also dependent on the density of the crowd.

We implement the experiments for maximum hops Hmax=1,2,3. Note that, the bracelet will direct

people to stop walking when it receives unsafe signal, and the density tends to be more evenly distributed in the entire region. To estimate the effectiveness of the configurations, we present the number of dense cells in the observed area and focus on the density improvement over time, and then discuss about energy consumed among these configurations.

(1) With and without secondary source propagation

Note that the Nmsg does not only determine whether the bracelet will become a message source,

but also decide how the bracelet directs its master, i.e.,to stop or to move. Considering that the later effect can influence the density of the area, we firstly compare the density with and without the function of secondary sources with the same Hmax=2 in a tunnel scenario with low

pedestrian-generated rate.

Figure 4.5a reflects that without the proposed transmission propagation, a smaller Nmsg can more

or less obtain lower density of the crowds, however, the number of dense cells seems to be linearly increasing over time with a similar rate. Figure 4.5b proves that the dense area can be amended efficiently by application of the way of generating new ”danger” sources according to the number of received messages. Even with Nmsg= 7, the increasing rate of density can be visibly decreased.

With this secondary source, the number of dense cells can be controlled to nearly constant given Nmsg=3. Based on this result, we will investigate an optimal configuration of the device with

secondary source in the later experiments. (2) Various Nmsg with secondary source

Given lower Nmsg, which is less stricter threshold to determine whether the pedestrian should stop

walking and the device should become a new source, we assume the density will decrease, while the power consumption can increases or decrease determined by the density and the threshold Nmsg

for secondary source generation.

Figure 4.6(left) demonstrates the number of dense cells over time when Nmsg varies from 2 to

8 given Hmax = 2 for all three scenarios. As expected, given lower Nmsg, the number of dense

cells increases with lower rate. Given Nmsg <4, the number of dense cells increases gently.

Fig-ure 4.6(right) presents the trend of power consumption and density of the area given various pre-determined Nmsg. We can observe that the power consumption does not decrease monotonically.

The detailed power consumption for each configuration can be found in Figure A.8, Figure A.11 and Figure A.14 for tunnel scenario with two pedestrians-entering rate and exit scenario respec-tively.

Figure 4.6a(left) shows that in the first scenario there appears dense cells at about 60s , and in one minute (at 120s), the number of dense cells does not increase greatly given Nmsg < 5 even

(23)

(a) Without secondary source

(b) With secondary source

Figure 4.5: Comparison of number of dense cells over time with and without secondary source given Hmax = 2 in tunnel scenario.

in this closed-tunnel structure, which allows the dangerous situation to be cleaned up sufficiently. Considering the power consumption, we can find an ideal Nmsg to be 4 or 5 in Figure 4.6a(right),

with which minimum energy is consumed and enough reaction time (less than 4 dense cells are generated in 1 minutes) is assured in this scenario. As shown in Figure A.8, the power consump-tion for propagaconsump-tion of primary sources(blue line) proves that, given a higher Nmsg, the density

increases, which causes more primary sources to be generated. And the energy consumed by the secondary sources is determined by Nmsg and the number of generated primary sources. With the

tradeoff of these two features, it does not increase or decrease monotonically, neither does the sum power consumption. And the dot line indicates how far a ”density message” can be propagated, furthermore, how it can be improved given smaller Nmsg. Given Nmsg=3, almost all the devices

have been committed messages relay, even those standing far away from the entrance of the tun-nel. As increasing Nmsg, the distance the message can reach become shorter and shorter. The

higher maximum power consumption with lower average value also demonstrates that overcrowded is formed in some specific small area instead of evenly distribution.

For the tunnel scenario with high-frequency generating agent, Figure 4.6b(left) presents from 50s to 75s, given Nmsg <=5, less than 2 dense cells are generated during this time, which gives

rela-tively enough time to react to a dangerous situation compared with greater Nmsg. Additionally,

taken power consumption into account, we can conclude 5 is an optimal configuration. (See Fig-ure 4.6b(right)).

(24)

Fig-ure 4.6c, 3 or 4 gives best performance on high density prevention with low power consumed. More details on power consumption for these two scenarios can be found in Figure A.11 and Figure A.14. From the above results, we obtain a similar optimal configuration, i.e., Nmsg=4, which is usually

adequate to control the density of crowded area. Different performance on power consumption given various Nmsg is presented, and it is not clear to conclude an overview of how Nmsg

deter-mines the power consumption. However, we can observe some common phenomenon that, given Nmsgwhich ensures sufficient density control, namely, a status of low and gently increased density

(Nmsg <5 for scenario1, Nmsg< 6 for scenario2, Nmsg <4 for scenario3), greater Nmsg consumes

lower energy. It reflects the Nmsg notably affects the formation of secondary sources in this case.

(3) Various Hmax with secondary source

Apparently, Hmax will offer different effectiveness for information awareness. For tunnel scenario

with low-frequency generating agents, as shown in Figure 4.7a(left), given Hmax=1, lower Nmsg

does not make significant improvement when the strategy of secondary source is applied, despite Nmsg=2 optimizes the area with less dangerous cells. We can see that the high density is

for-mulated at around time 55s. Given Nmsg=2, about 10 dense cells of 1m2 area are formed in 1

minute, and the density continues increasing with a similar rate. This does not achieve a sufficient early-warning function. From Figure 4.7b(left) and Figure 4.7c(left), we can witness a noticeable improvement on density control. Given Hmax = 2 with Nmsg <= 5 or Hmax = 3, the number of

dense cells in 1 minute stay under around 4, which is considered more applicable to manage the congestion.

The above results demonstrates the satisfying effectiveness of Hmax and Nmsg with respect to

safety in this scenario, and we are also interested in the energy consumption under these thresh-olds. As shown in Figure 4.7a(right), the power consumption is almost increasing with the same rate as the density. Detailed power consumption over time for Hmax=1 can be found in

Fig-ure A.7. It is shown that, when Nmsg < 4, the sum of power consumed depends mainly on the

one from primary sources, i.e. the surrounding density of the device. This is due to the secondary sources are generated based on the primary one, and the small Hmax does not help to generate

more secondary sources and consume more power. As Nmsg grows, the effect of this threshold on

secondary sources becomes visible.(See Figure A.7d, A.7e, A.7f, A.7g) Therefore, the power con-sumption for secondary sources decreases while even more primary sources are formulated. The fraction of power consumed by secondary sources in sum of power decreases, and the primary one accordingly increases. From Figure 4.7b(right), we can notice that, given Hmax=2, the power

consumption decreases under a low density(average number of dense cells < 2) and then increases under a higher density, while the density keeps growing as the Nmsg increases. This indicates that

Nmsg makes more detectable influence under low density than high density on generation of the

secondary sources. Given Hmax=3, the average number of dense cells is kept below 2, so various

Nmsg makes little difference on the number of primary sources. Therefore, the increasing Nmsg

under low density will noticeably limit the generation of secondary sources (See Figure 4.7c(right)). Detailed power consumption given Hmax=2 and 3 with various Nmsg are presented in Figure A.8

and Figure A.9 respectively.

Given Hmsg=2, we can keep average no more than 2 dense cells with Nmsg <5, and the power

consumption is less than 10,000,000 nj in this experiment.(See Figure 4.7b(right)) Hmax= 3 offers

the same density, however, cost double energy(> 20,000,000 nj). The difference of energy con-sumed between these two Hmaxcan be even greater if the message size is greater or during longer

processing period. Therefore, Hmax=2 is a better option for in this case with regard to power

consumption as well as safety.

From the same perspective, we obtain a similar conclusion in the other two scenarios (See Fig-ure 4.8 and FigFig-ure 4.9). Hmax=2 with some optimal options of Nmsg, 4 or 5 in these scenarios,

provides ideal performance with regards to both density and power consumption. Detailed power consumption for each configuration in tunnel scenario with high-frequency generating agents are presented in Figure A.10, Figure A.11 and Figure A.12. Same figures for exit scenario can be found

(25)

in Figure A.13, Figure A.14 and Figure A.15. Because the device stops consuming energy when it has left from the exit of the square, power consumption for each devices varies and presents shaking distribution in these figures, which don’t appear in two other scenarios.

To demonstrate the effect of the usage of this device, Figure 4.10 presents the density distribution for exit scenario applying 3 device prototypes. As shown in the pictures, only applying the variable sensing range will only increase the network connectivity, but not prevent the congestion at the exit. Secondary sources help to slow down the pedestrians herding into the exit, so that develop a evenly distributed crowd.

(26)

(a) Tunnel scenario with low-frequency generating agents

(b) Tunnel scenario with high-frequency generating agents

(c) Exit scenario

Figure 4.6: The number of dense cells and consumed power for various Mmsg given Hmax=2. Left:

(27)

(a) Hmax=1

(b) Hmax=2

(c) Hmax=3

Figure 4.7: Number of dense cells and power consumption for tunnel scenario with low-frequency generating agents. Left: Number of dense cells over time for various Nmsg given Hmax = 1,2,3;

Right: Average number of dense cells and average power consumption for various Nmsg given

(28)

(a) Hmax=1

(b) Hmax=2

(c) Hmax=3

Figure 4.8: Number of dense cells and power consumption for tunnel scenario with high-frequency generating agents. Left: Number of dense cells over time for various Nmsg given Hmax = 1,2,3;

Right: Average number of dense cells and average power consumption for various Nmsg given

(29)

(a) Hmax=1

(b) Hmax=2

(c) Hmax=3

Figure 4.9: Number of dense cells and power consumption for exit scenario. Left: Number of dense cells over time for various Nmsg given Hmax = 1,2,3; Right: Average number of dense cells and

(30)

(a) Fixed communication range

(b) Variable communication range without secondary sources

(c) Variable communication range with secondary sources

Figure 4.10: Density distribution on exit scenarios given 3 configuration, nodes with blue circle is included in giant component, and radius of circle reflects the communication range.

(31)

Chapter

5

Conclusions and Future Works

5.1 Conclusions

In this thesis, we propose a simple and cheap wireless bracelet, which aims to prevent high density crowds by density detection, information propagation and instruction of pedestrians’ movement in area of interest. Variable communication range and secondary source are applied with the as-sumptions that early warning signs are valuable and even vital in stampedes prevention, especially for a tunnel or queue scenario. Experiments on three scenarios shows the dense regions can be significantly reduces under the instruction of this device, and the performance with regard to power consumption is improved with two proposed features.

In the social force model, keeping a bracelet sensing 5 to 8 neighbors can mostly obtain a fully-connected network for all three scenarios. It suggests that the area structure and entering frequency of pedestrian flows do not materially influence the optimized design of sensing range threshold [Ninc, Ndec], because the sensing range is adapted to only its current density over time. It is

in-teresting to see that a greater communication range and a lower transmission threshold will not always cause higher consumption. Because of the tradeoff between the density and the danger message threshold Nmsg, there exits an optimal option for Nmsg on each scenario. In the

exper-iments, Nmsg=4 gives ideal results for all the three scenarios. Instead of a fixed greater Hmax,

whose value is usually adopted according to the specific area structure, we prove that Hmax=2

with secondary sources allows the danger messages are generated and propagated mainly based on the density around the device.

This device model provides an simple prototype for crowd management, which functions according to merely local density. Thanks to its simplicity and density-oriented, we do not need to pay much effort on analysis of the structure of the zone or the pedestrians motion model when designing the configuration of this device. Furthermore, its self-organized effect makes individual instruction for its wearer to avoid walking to the danger based on its spatiotemporal situation. Therefore, it provides a flexible and easily implemented device to control and manage crowds in large gathering events.

5.2 Future Works

In this model, we assume the less danger cell there is, the more security the region is. However, it is still important to define the safety standard, namely, the tolerated maximum number of danger cells for each scenario and an sufficient reaction time to dangerous situation. Given these informa-tions, the device can be configured to an optimal working status.

(32)

The proposed strategy of secondary sources decreases the density of the crowds, however, it fails to indicate the approximate distance from the real density location. This can be improved by some distinguishing ways, such as adding an additional bit of data to point out the message sort, and providing an yellow light to differentiate the signal type. Additionally, in this model, we do not consider the transmission delay and the network interruption, which is also important in practical applications. Providing each person on device may also not be realistic under extremely large gathering events, due to the economic concerns and the possibility of networks transmission congestion. Thus, an optimization method to reduce the necessary number of devices is required. For example, in crowded area, pedestrians locating close can be considered in the similar condition. This can be taken into account to reduce the information redundancy to optimize the transmission model. Spatiotemporal density information can also coordinate with data from other sensors, such as camera, to produce more concrete and comprehensive output of the crowds states.

(33)

Bibliography

[1] U.S. Census Bureau. U.s. and world population clock, 2015. [2] WHO. Urban population growth, 2015.

[3] Wikipedia. List of human stampedes — wikipedia, the free encyclopedia, 2015. [Online; accessed 15-June-2015].

[4] Wikipedia. 2014 shanghai stampede — wikipedia, the free encyclopedia, 2015. [Online; accessed 15-June-2015].

[5] Wikipedia. Love parade — wikipedia, the free encyclopedia, 2015. [Online; accessed 15-June-2015].

[6] Wikipedia. 2013 kumbh mela stampede — wikipedia, the free encyclopedia, 2015. [Online; accessed 15-June-2015].

[7] Wikipedia. 2013 madhya pradesh stampede — wikipedia, the free encyclopedia, 2015. [Online; accessed 15-June-2015].

[8] Teo Yu Hui Angela, Vaisagh Viswanathan, Michael Lees, and Wentong Cai. Analysing the effectiveness of wearable wireless sensors in controlling crowd disasters. Procedia Computer Science, 29:1590–1599, 2014.

[9] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995.

[10] Dirk Helbing. Social self-organization: Agent-based simulations and experiments to study emergent social behavior. Springer, 2012.

[11] Suiping Zhou, Dan Chen, Wentong Cai, Linbo Luo, Malcolm Yoke Hean Low, Feng Tian, Vic-tor Su-Han Tay, Darren Wee Sze Ong, and Benjamin D Hamilton. Crowd modeling and sim-ulation technologies. ACM Transactions on Modeling and Computer Simsim-ulation (TOMACS), 20(4):20, 2010.

[12] Minoru Fukui and Yoshihiro Ishibashi. Self-organized phase transitions in cellular automaton models for pedestrians. Journal of the Physical Society of Japan, 68(8):2861–2863, 1999. [13] Dietmar Bauer. Comparing pedestrian movement simulation models for a crossing area based

on real world data. In Pedestrian and Evacuation Dynamics, pages 547–556. Springer, 2011. [14] Soraia Raupp Musse and Daniel Thalmann. Hierarchical model for real time simulation of

vir-tual human crowds. Visualization and Computer Graphics, IEEE Transactions on, 7(2):152– 164, 2001.

[15] Mehdi Moussa¨ıd, Dirk Helbing, and Guy Theraulaz. How simple rules determine pedestrian behavior and crowd disasters. Proceedings of the National Academy of Sciences, 108(17):6884– 6888, 2011.

(34)

[16] Dirk Helbing and Pratik Mukerji. Crowd disasters as systemic failures: analysis of the love parade disaster. EPJ Data Science, 1(1):1–40, 2012.

[17] Dirk Helbing. Models for pedestrian behavior. arXiv preprint cond-mat/9805089, 1998. [18] Dirk Helbing, Lubos Buzna, Anders Johansson, and Torsten Werner. Self-organized pedestrian

crowd dynamics: Experiments, simulations, and design solutions. Transportation science, 39(1):1–24, 2005.

[19] Mikel Rodriguez, Ivan Laptev, Josef Sivic, and Jean-Yves Audibert. Density-aware person detection and tracking in crowds. In Computer Vision (ICCV), 2011 IEEE International Conference on, pages 2423–2430. IEEE, 2011.

[20] Wouter G Van Toll, Atlas F Cook, and Roland Geraerts. Real-time density-based crowd simulation. Computer Animation and Virtual Worlds, 23(1):59–69, 2012.

[21] Jinghui Zhong, Nan Hu, Wentong Cai, Michael Lees, and Linbo Luo. Density-based evolu-tionary framework for crowd model calibration. Journal of Computational Science, 6:11–22, 2015.

[22] Ernesto L Andrade, Scott Blunsden, and Robert B Fisher. Hidden markov models for op-tical flow analysis in crowds. In Pattern Recognition, 2006. ICPR 2006. 18th International Conference on, volume 1, pages 460–463. IEEE, 2006.

[23] Ramin Mehran, Akira Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 935–942. IEEE, 2009.

[24] M Nakatsuka, H Iwatani, and J Katto. A study on passive crowd density estimation using wireless sensors. In The 4th Intl. Conf. on Mobile Computing and Ubiquitous Networking (ICMU 2008), 2008.

[25] Yaoxuan Yuan, Chen Qiu, Wei Xi, and Jizhong Zhao. Crowd density estimation using wireless sensor networks. In Mobile Ad-hoc and Sensor Networks (MSN), 2011 Seventh International Conference on, pages 138–145. IEEE, 2011.

[26] Yaoxuan Yuan, Jizhong Zhao, Chen Qiu, and Wei Xi. Estimating crowd density in an rf-based dynamic environment. Sensors Journal, IEEE, 13(10):3837–3845, 2013.

[27] Shaogang Gong and Tao Xiang. Recognition of group activities using dynamic probabilistic networks. In Computer Vision, 2003. Proceedings. Ninth IEEE International Conference on, pages 742–749. IEEE, 2003.

[28] Dan Kong, Douglas Gray, and Hai Tao. Counting pedestrians in crowds using viewpoint invariant training. In BMVC. Citeseer, 2005.

[29] Nuria Oliver, Ashutosh Garg, and Eric Horvitz. Layered representations for learning and inferring office activity from multiple sensory channels. Computer Vision and Image Under-standing, 96(2):163–180, 2004.

[30] Alois Ferscha and Kashif Zia. Lifebelt: Silent directional guidance for crowd evacuation. In Wearable Computers, 2009. ISWC’09. International Symposium on, pages 19–26. IEEE, 2009. [31] Daniel Roggen, Martin Wirz, Gerhard Tr¨oster, and Dirk Helbing. Recognition of crowd behavior from mobile sensors with pattern analysis and graph clustering methods. arXiv preprint arXiv:1109.1664, 2011.

[32] Niwat Thepvilojanapong, Shin’ichi Konomi, Yoshito Tobe, Yoshikatsu Ohta, Masayuki Iwai, and Kaoru Sezaki. Opportunistic collaboration in participatory sensing environments. In Proceedings of the fifth ACM international workshop on Mobility in the evolving internet architecture, pages 39–44. ACM, 2010.

(35)

[33] Constantine K Christakos. Sensor networks applied to the problem of building evacuation: An evaluation in simulation. Proc. 15th IST Mobile and Wireless Summit, pages 134–138, 2006.

[34] Wendi Rabiner Heinzelman, Anantha Chandrakasan, and Hari Balakrishnan. Energy-efficient communication protocol for wireless microsensor networks. In System sciences, 2000. Pro-ceedings of the 33rd annual Hawaii international conference on, pages 10–pp. IEEE, 2000. [35] Wail Mardini, Yaser Khamayseh, and AL-Eide Shorouq. Optimal number of relays in

coop-erative communication in wireless sensor networks. 2012. [36] Sean Luke. Multiagent simulation and the mason library.

[37] DAVID Barron and XANTH ´E Mallett. The hillsborough football stadium disaster, april 15, 1989. Disaster Victim Identification: Experience and Practice, 76:109, 2011.

[38] Sindhu Kolli and Kamalakar Karlapalem. Mama: multi-agent management of crowds to avoid stampedes in long queues. In Proceedings of the 2013 international conference on Autonomous agents and multi-agent systems, pages 1203–1204. International Foundation for Autonomous Agents and Multiagent Systems, 2013.

(36)

(37)

Appendix

A

(38)

(a) Fixed Communication Range (b) Ninc=3, Ndec=6

(c) Ninc=4, Ndec=7 (d) Ninc=5, Ndec=8

(e) Ninc=6, Ndec=9 (f) Ninc=7, Ndec=10

(39)

(40)

(g) Ninc=8, Ndec=11 (h) Ninc=9, Ndec=12

(41)

(42)

(43)

(44)

(a) Hmax= 1, Nmsg= 2 (b) Hmax= 1, Nmsg= 3

(c) Hmax= 1, Nmsg = 4 (d) Hmax= 1, Nmsg= 5

(e) Hmax= 1, Nmsg = 6 (f) Hmax= 1, Nmsg= 7

(g) Hmax= 1, Nmsg = 8

Figure A.7: Tunnel scenario with low-frequency generating agents: power consumption distribution for various Nmsg given Hmax=1. The dot lines indicate how far the messages can be transmitted.

Devices after the dot line do not consume energy, which means the message are not propagated that far.

(45)

(a) Hmax= 2, Nmsg = 2 (b) Hmax= 2, Nmsg = 3

(c) Hmax= 2, Nmsg = 4 (d) Hmax= 2, Nmsg = 5

(e) Hmax= 2, Nmsg = 6 (f) Hmax= 2, Nmsg = 7

(46)

(47)

Figure A.10: Tunnel scenario with high-frequency generating agents: power consumption distri-bution for various Nmsg given Hmax=1. The dot lines indicate how far the messages can be

transmitted. Devices after the dot line do not consume energy, which means the message are not propagated that far.

(48)

(49)

A Wireless Bracelet: Self-Organization of Pedestrians for High Density Prevention in Crowds