Simulating heterogeneous crowds from a physiological perspective

Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Contents lists available at ScienceDirect

Neurocomputing journal homepage: www.elsevier.com/locate/neucom

Simulating heterogeneous crowds from a physiological perspective Liping Zheng n, Dang Qin, Yajun Cheng, Lin Wang, Lin Li School of Computer and Information, Hefei University of Technology, China

art ic l e i nf o

a b s t r a c t

Article history: Received 31 October 2013 Received in revised form 10 December 2014 Accepted 13 December 2014

Most of the existing approaches to simulate heterogeneous crowd behaviors focus on the aspect of psychology. From a human's physiological characteristics perspective, this paper presents a method to generate heterogeneous crowd behaviors. We choose four basic physiological characteristics, including gender, age, health and body shape, and map them to a navigation method, which is reciprocal velocity obstacle approach in the paper. The mapping parameters are determined through a two-step process. Firstly, a video based method is proposed to obtain simulation parameters for single physiological feature by tracking and analyzing trajectories of persons in real video scenes, and then a comprehensive mapping is presented to combine all characteristics together to generate parameters for a certain person. Through a number of simulations and validation experiments, we demonstrate that the proposed method is simple but effective and efficient to exhibit heterogeneous behaviors of crowds. & 2015 Elsevier B.V. All rights reserved.

Keywords: Crowd simulation Heterogeneous behaviors Physiological characteristics Video tracking

1. Introduction Crowd simulation has been extensively studied and applied to many fields including movies, games, and virtual simulations. A virtual crowd, formed by people with different and various appearances and behaviors, is the so-called heterogeneous, which is natural in real world and important for many applications. This paper focuses on the aspect of crowd behavior heterogeneity, and most of the existing works utilize psychological characteristics and personality models to produce individual differences. For instance, Guy et al. [1] simulated heterogeneous crowd using personality trait theory based on reciprocal velocity obstacle (RVO) library [2], where a series of user study experiments were done to derive a linear mapping from personality descriptors to RVO parameters to control the extent that agents show various behaviors. To the best of our knowledge, little effort is devoted to utilizing physiological characteristics to generate behavior heterogeneity. Definitely, this does not mean, however, that physiological features are trivial for simulating a crowd. We think that many works focus on psychology aspect because it is easier to implement. There are many mature and recognized models for personality, such as PEN (Psychoticism, Extraversion, and Neuroticism) and OCEAN (Openness , Conscientiousness, Extraversion, Agreeableness, and Neuroticism), and these models categorize personality into several discrete and orthogonal types. Nevertheless, the physiological features are quite a lot and are mostly continuous, such as age,

n

Corresponding author. Tel.: +86 551 62901377. E-mail address: [email protected] (L. Zheng).

health, physical power, weight and height, so they are hard to model and lack mature theories about their effects on behaviors. However, simulating behaviors from physiological perspective are more natural and instinctive, for instance, the velocity and radius of occupied space of a person are more relevant to physiological aspects than others including psychology. Hence, we try to simulate heterogeneous crowd from a pure physiological perspective. Inspired by the work of [1], we also choose RVO library as our simulation algorithm, however, it can be easily replaced by other approaches. Four basic physiological characteristics, including gender, age, health and body shape, are taken into account. We determine a mapping from a single physiological feature to RVO parameters through video tracking based method, and a comprehensive mapping, which combines these four characteristics together, is proposed to generate various simulation parameters for agents exhibiting heterogeneous behaviors. The main contribution of this paper is an anthropometric perspective to simulate heterogeneous behaviors of a crowd. This is the first attempt to conduct crowd simulation completely from physiology aspect. Extended from [3] but instead of using user studies and statistical method, this paper presents a video based analysis approach, which is more direct and objective, to analyze persons in the videos and capture their trajectories in order to determine the simulation parameters. Additionally, relative values of these parameters are used to reduce computing errors on the experimental analysis stage. Experiments show that the proposed approach is simple, light-weight yet effective to generate visually convincing crowd simulation results with excellent heterogeneity. The rest of the paper is organized as follows. Related works in crowd simulation and behavior difference modeling are described

http://dx.doi.org/10.1016/j.neucom.2014.12.103 0925-2312/& 2015 Elsevier B.V. All rights reserved.

Please cite this article as: L. Zheng, et al., Simulating heterogeneous crowds from a physiological perspective, Neurocomputing (2015), http://dx.doi.org/10.1016/j.neucom.2014.12.103i

L. Zheng et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

2

in Section 2. In Section 3, a brief introduction to RVO parameters is given. And we highlight the proposed mapping methods in Section 4. Section 5 demonstrates both simulation results and validation experiments.

2. Related works 2.1. Heterogeneous crowd

Parameters timeHorizon and timeHorizonObst are the minimum amount of time for which the agent's speed, computed by the simulation, is safe with respect to other agents and obstacles, respectively. Here, we call them together as planning horizon. The larger the planning horizon parameter, the sooner this agent will respond to the presence of other agents or obstacles, that is, the agent is more foresight and of longer vision. Parameters neighborDist and maxNeighbors represent the maximal distance and number, respectively, of other agents that the agent takes into account in the path planning. We call these two parameters together as planning scope. Basically, the larger the planning scope parameter, the more consideration of other agents and the finer of the simulation, of course, with longer calculating time and cost. Parameter prefVelocity is the preferred velocity the agent would take if there are not any other agent or obstacle around. RVO library will achieve a tradeoff between this speed and that will guarantee no collisions. Parameter radius is the radius of occupied space of an agent. This hard constraint can be obtained from physiological body feathers of an agent, as well as psychological aspects [1].

There are several mature algorithms and models for crowd simulation, such as classical Boids [4] and social force method [5], and recent reciprocal velocity obstacles (RVO) algorithm [2], fluid based models [6] and data driven approaches [7]. We choose RVO as the navigation library. The heterogeneity aspect of a crowd is the focus of this paper. Existing works produce heterogeneous crowds roughly from three aspects: individual model, controlled clustering and autonomous behavior. Crowd simulation credibility will benefit a lot from variety of textured appearances [8,9], as well as body shapes and poses [10]. At the same time, controlled clustering can explore some special distributions [11] and formations [12], and is extensively studied in the multi-robot control field. The following will focus on behavior difference of human crowd.

4. Video-analysis based mapping method

2.2. Behavior diversification

4.1. Mapping physiological characteristics

Jentsch et al. [13] conducted a comprehensive literature review to summarize the social and psychological individual-difference works. As a result, they found that academia concerned relatively little about individual differences from anthropometric and psychological aspects. However, some existing literatures make use of internal characteristics of a person to model behaviors. Personality has been adopted for generating various behaviors. Guy et al. [1] used PEN based personality trait theory to drive agents exhibiting complex variations in behaviors. Similarly, Durupinar et al. [14,15] utilized OCEAN personality model to simulate crowd behaviors by mapping from personality traits to existing behavior types driven by Hi-DAC. Curtis et al. [16] tried to simulate heterogenous individuals with age and gender, but only simply attached them to the agents' preferred speed and maximum speed. Another way to achieve heterogeneity goal is cloning behavior from existing video samples. Copy and paste technique is used to produce heterogeneous crowd motion from different sources and patterns [7,17,18]. These methods are effective to generate good results, however, are restricted to the quality and quantity of video samples. As to physiological characteristics, Kaup et al. [19] proposed age-based crowd behavior simulation by modeling age differences as the strength of forces through social force method. Pelechano et al. [20] presented Hi-DAC model, which assigned agents with different psychological and physiological traits for generating individual-different behaviors. This paper presents a physiology perspective to produce the diversification of a crowd. Existing psychology based method inspires this work, and video analysis based method will be brought into our paper for determining simulation parameters.

The main object of this paper is to simulate a heterogeneous crowd by leveraging physiological characteristics. Since RVO library is powerful and easy to use, we choose it as a carrying navigation method to render the differences between each individual. Hence, mapping from physiology features to simulation parameters must be considered at first. Physiological characteristics are inherent to human, and have measurable impacts on people's behavioral reactions. As to gender, Conner [21] presented that women had a more sensitivity and an enhanced physical alarm response than men when facing to danger or threat, however, most of men had own a better overall situation than women. Lobjois and Cavallo [22] studied age-related effects on street-crossing decisions, and the results showed that older people tended to choose a greater mean time gap between vehicles to compensate for their increased crossing time, so their crossings are sooner than the younger ones. This indicates that age difference influences their judgments about collision avoidance in walking or running. At the same time, it is obvious that human's age and body shape are directly related to the occupied space, and health is related to their walking speed [23]. Now we need to present a qualitative mapping from person's main four physiological characteristics to the RVO's six parameters. Theoretically, each physiological feature correlates with each RVO parameter. However, for the sake of simplification and operability, we consider only the principal factors and ignore the minor relations. According to the above analysis, males have broad overall situation to consider further distance and more neighbors, but have less agile response on collision avoidance than females, so gender is mapped to neighborDist, maxNeighbors, timeHorizon and timeHorizonObst. And age is the same as gender for the reason that maturity is directly related to the behavior of interaction with each other or obstacles. People's age and body shape determine the occupied space, so we associate them with radius. Also, weak or strong people have different walking speed, so health condition is solely related to prefVelocity. The overall mapping is depicted in Fig. 1. Note that planning scope and planning horizon are used here so the six parameters are simplified to only four. Based on Fig. 1, the following section will define the quantitative mapping.

3. RVO library and parameters RVO library [2] presents a formal approach to perform collision avoidance among reciprocal agents, and each agent acts independently and cannot communicate with others. It provides a set of interface parameters, including timeHorizon, timeHorizonObst, neighborDist, maxNeighbors, prefVelocity and radius.

Please cite this article as: L. Zheng, et al., Simulating heterogeneous crowds from a physiological perspective, Neurocomputing (2015), http://dx.doi.org/10.1016/j.neucom.2014.12.103i

L. Zheng et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

4.2. Mapping parameters determination

3

Table 2 RVO parameters and their default values.

After the qualitative mapping between physiological characteristics and simulation parameters is constructed, quantitative relationship needs to be further defined by determined, which is conducted by a four-step process as follows. 4.2.1. Fuzzy discrete level decomposition Firstly, in order to determine the quantitative mapping parameters, each of the four human's physiological characteristics is discretely classified into three fuzzy levels, which is depicted in Table 1. Here “middle-aged”, “asexual”, “normal body shape” and “normal health” are moderated levels, and are thus set as the default parameters in Table 2. They serve as horizon standards to relatively tune the mapping parameters. In real life, “asexual” does not exist but is still needed for comparison. Even though the classification is simple and rough, we can obtain up to 54 samples of human from physiology aspect than that of psychology proposed in [1], which is only 9. This decomposition can enrich the diversity and heterogeneity of crowd simulation. An example of these samples could be a person who is a thin and weak old woman. According to RVO library and results of [1] , the default values of these parameters are given in Table 2. These default values are assigned to a person who belongs to all the middle levels in Table 1. Based on default values, we only need to generate a relative quantity for each fuzzy level. 4.2.2. Video analysis method In this paper, we present a video analysis method to determine simulation parameters for planning horizon, planning scope and prefVelocity, while radius is handled by user study method and to be discussed in the next section. The procedure of the proposed method is depicted in Fig. 2 and will be described as follows. Scenario choosing and video capturing: We have taken about one-hour video samples in a busy street. The videos are captured

Physiological characteristics

RVO parameters

Gender

planning horizon

Parameter

Default value

timeHorizon timeHorizonObst neighborDist maxNeighbors prefVelocity radius

10.0 s 10.0 s 15.0 m 10 1.45 m/s 1.0 m

Compressive tracking

Scenario choosing video capturing

Mapping parameters determination excluding radius

Referents selecting

Object tracking

Person sampling

Trajectories

prefVelocity Calculating planning scope

planning horizon

Fig. 2. The framework of the proposed video analysis based method. The dotted lines mean the correlation relationship and will be described in Section 4.2.2.

timeHorizon timeHorizonObst

Age

planning scope neighborDist maxNeighbors

Health

prefVelocity

Body shape

radius

Fig. 3. Scenario of the captured videos. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this article.)

Fig. 1. Mapping from four physiological characteristics to six RVO parameters. The dotted lines mean the correlation relationship and will be described in Section 4.2.2.

Table 1 A fuzzy classification for four physiological characteristics. Gender

Age

Health

Body shape

Male Asexual Female

Child Middle-aged Old-aged

Weak Normal Strong

Thin Normal Fat

with a fixed viewpoint (451), a fixed frame rate (25 fps) and the resolution is 720n576. There are a few environmental obstacles in the scenario, and they are marked as red rectangles in Fig. 3. These obstacles are interlacedly placed so that most of pedestrians should perform obstacle avoidance. We pay attention to persons walking horizontally across the scenario, thus the perspective distortion of camera needs not much consideration. Pedestrians or vehicles moving vertically in the scene are treated as dynamic obstacles. They are marked as blue rectangles in Fig. 3. Of course, the oppositely walking persons are reciprocal obstacles to each other. Referents selecting: We elaborately select more than 100 tracking referents following the rules: first, when simulation parameters of one physiological feature are determined, all other characteristics are selected as the middle levels of Table 1; second, typical samples are chosen as best possible. For an example, when

Please cite this article as: L. Zheng, et al., Simulating heterogeneous crowds from a physiological perspective, Neurocomputing (2015), http://dx.doi.org/10.1016/j.neucom.2014.12.103i

L. Zheng et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

4

Fig. 4. Captured different trajectories by compressive tracking algorithm. The left is a man and the right is a woman. Trajectories are plotted as red curves. It can be clearly seen that the female responds to obstacles earlier. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this article.)

analyzing parameters for gender, we select samples with nearly the same age, health and body shape. Also, according to age analysis, we choose typical old, middle persons and children referents with 70, 25 and 5 years old, respectively, which are distinct for better heterogeneity. Trajectories tracking: Recently, there are many image processing [24,25], video analysis and tracking methods [26–29]. Here we use compressive tracking algorithm proposed by Zhang et al. [26] for object tracking because it is robust, simple yet effective. The algorithm needs an interactive initial to choose objects to be tracked, and then generates positions of the tracked object every frame, denoted by pj at frame j. The trajectory T ri for referent ri within n frames is denoted as T ri ¼ fp1 ; …; pk ; …; pn g

Trajectory Obstacle

Fig. 5. Processing of the trajectory generated by compressive tracking algorithm. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this article.)

for k ¼ 1; 2; …n

Once T ri of a person is obtained, we can analyze the behaviors. Fig. 4 visualizes the trajectories of a young male and a young female, and they exhibit distinguished planning horizon time while avoiding obstacles. Parameters calculating: We can compute the average velocity V ri and the planning horizon time Phri for ri according to generated trajectories T ri . Obviously, V ri , to be assigned to parameter prefVelocity, can be obtained by integral method as  n 1  X J pk þ 1  pk J f where f ¼ 25 fps V ri ¼ n k¼1 In most of our experimental scenarios, the observed trajectories have two inflection points. One is at the place starting to plan the path avoiding the obstacles, and the other is where the referent is being bypassing obstacles. These two points are denoted as yellow dots in Fig. 5. In addition, T ri is very noisy. For these reasons, we fit the curve with a cubic function, resulting in T 0ri , plotted as the blue curve in Fig. 5. Then, the first turn point can be easily determined with its frame number as Fst, and the latter as Fend. In some situations, the latter point does not exist hence the cubic can degrade to quadratic. If so, it can be found by projecting vertically the obstacle position to the curve. Finally, Phri , to be assigned to parameter planning horizon, can be computed by Phri ¼

Inflection points

F end  F st f

Note that in order to reduce calculating errors, here we need to compute the relative quantities at first, then obtain the absolute value on the basis of the default values of Table 2. Taking prefVelocity as an example, we get values Vnormal, Vweak and Vstrong for normal, weak and strong referents by video based method,

then relative values of them, denoted by V 0normal , V 0weak and V 0strong , are computed by V 0normal ¼ 0 V  V normal V 0weak ¼ weak V normal V strong  V normal V 0strong ¼ V normal

ð1Þ 0

Similarly, planning horizon parameters for age, that is Phchild , 0 and Phold  aged , can be determined through Phchild, Phmiddle-aged and Phold-aged. However, it is special for gender since 0 0 there is no ”asexual” people, and Phmale and Phfemale can be computed from Phmale and Phfemale by 0 Phmiddleaged

0

Phasexual ¼ 0 0

Phmale ¼ 0

Phmale  Phfemale Phmale þ Phfemale

Phfemale ¼

Phfemale  Phmale Phmale þ Phfemale

ð2Þ

Parameter planning scope: It is hard to determine planning scope since it may not be observed explicitly in the experiments. The results from [1] by using user study method are somewhat confusing. Hence, instead of ascertaining directly, we attach this parameter to planning horizon, that is to say, planning scope is changed relatively keeping pace with planning horizon. This correlation makes sense because a person, who tends to conduct a long-term path planning in time, will correspondingly consider more neighbors and longer distance. In other words, there is a positive relation between them. We mark this correlation as dotted lines in both Figs. 1 and 2. Formally, the planning scope

Please cite this article as: L. Zheng, et al., Simulating heterogeneous crowds from a physiological perspective, Neurocomputing (2015), http://dx.doi.org/10.1016/j.neucom.2014.12.103i

L. Zheng et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

parameter for age, denoted by Ps0child , Ps0middleaged and Ps0oldaged , is computed by

0

0

Gender--> Planning horizon sample

12

Ps0middleaged ¼ 0 Ps0child ¼ Phchild ¼

5

10

Phchild  Phmiddleaged Phmiddleaged

Ps0oldaged ¼ Pholdaged ¼

Pholdaged  Phmiddleaged : Phmiddleaged

8

ð3Þ

6 4 2

4.2.3. Mapping parameters to radius Similar to planning scope, parameter radius is key to crowd simulation but difficult to define. Both age and body shape are its determining factors. Based on our user study results about bustwaist-hip data of around 100 people of different age and body shape, empirical data are given here. The relative parameters for child/old-aged people are  20%/þ 10% while thin/fat people are 40%/ þ60%, respectively. Denoted by R0C age ; R0C bodyshape , their values are as follows:

0

frame

0

30

60

90

120

Fig. 6. Results of mapping from gender to planning horizon. Red and yellow diamond dots are male and female samples, respectively. The red and yellow line segments are average values, and the blue one is the computed result for “asexual” level. The x-axis represents the frame numbers between the two inflection points of a trajectory and the y-axis represents samples. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this article.)

R0C age A f 20%; þ 0%; þ 10%g;

Age --> Planning horizon

where C age A fchild; middleaged; oldagedg

sample

12

R0C bodyshape A f  40%; þ 0%; þ 60%g; where C bodyshape A fthin; normal; fatg

10

ð4Þ

8 6

4.2.4. Mapping parameters combination The above section determines simulation parameters for single physiological characteristic. Now we need to combine all of them and calculate the actual values, which is a mapping CR : C ri -RVOri . The formula is as follows: C ri ¼ ½timeHorizon; timeHorizonObst; neighborDist; maxNeighbors; prefVelocity; radiusT RVOri ¼ RVOdefault ðE þ αC ri :gender þ βC ri :age þ γ C ri :health þ θC ri :bodyshape Þ

4 2 0

frame

0

20

40

60

80

100

120

140

160

Fig. 7. Results of mapping from age to planning horizon. Red, blue, and yellow dots/ line segments are single/average result for child, middle-aged and old-aged group, respectively. The diamond dots are for male, and triangle ones for female. The average results are 73, 90, and 116, respectively. The x-axis and the y-axis are the same as Fig. 6. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this article.)

ð5Þ where C ri , RVOri , and RVOdefault are all six dimensional column vectors, E is [1,1,1,1,1,1]T, and RVOdefault is defined according to Table 2. α, β, γ, θ are combination weights, and they need not to be normalized for salience consideration. However, we clamp all of them to certain value ranges given by [1] conforming to validness requirements, for instance, the walking speed of human is limited in the range of [1.2 m/s, 2.2 m/s]. In the following experiments, all combination weights are equally set as 1.0.

Health --> prefVelocity

sample

12 10 8 6 4

4.2.5. Results of mapping parameters Here we give the experiment results. Firstly, we determine planning horizon parameters for gender. We select 10 men and 10 women in videos, and they are all typical middle-aged healthy persons. The results are painted in Fig. 6. By average, we get the planning horizon times are 83 and 97 frames, for male and female 0 0 respectively. Then according to Formula (2), Phmale and Phfemale are 7.8% and þ7.8%, respectively. And planning scope parameters Ps0male and Ps0female are the same with those of planning horizon. Secondly, planning horizon parameters for age are obtained similarly. Three groups are chosen separately as children, middleaged and old-aged persons, and each of them is composed of 5 males and 5 females in order to eliminate the effect of gender. The results are depicted in Fig. 7. Computed by Formula (3), the 0 0 relative values of Phchild (Ps0child ) and Pholdaged (Ps0oldaged ) are 18.9% and þ28.9%, respectively. Thirdly, calculated by Formula (1), mapping results from health to prefVelocity, denoted by V 0weak and V 0strong , are  51.6% and þ38.7%, respectively. The results are demonstrated in Fig. 8.

2 0

pixel/frame

0

1

2

3

4

5

6

Fig. 8. Results of mapping from health to prefVelocity. Red, blue and yellow dots denote weak, normal and strong individuals, respectively. The average results, painted as line segments, are 1.5, 3.1, and 4.3, respectively. The x-axis represents average shifting pixels per frame and the y-axis represents samples. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this article.)

Based on the above planning horizon, planning scope and prefVelocity results, and radius values obtained by Formula (4), an overall table can be achieved for all fuzzy levels, depicted in Table 3. According to Table 3, we can obtain the specific parameters of a person by using Formula (5). For instance, simulation parameters, including relative and actual values, of a thin, weak and old-aged male can be easily figured out as Table 4.

Please cite this article as: L. Zheng, et al., Simulating heterogeneous crowds from a physiological perspective, Neurocomputing (2015), http://dx.doi.org/10.1016/j.neucom.2014.12.103i

L. Zheng et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

6

Table 3 Simulation parameters for physiological characteristics.

Physiology classification Gender Male Asexual Female Age Child Middle-aged Old-aged Health Weak Normal Strong Body shape Thin Normal Fat

planning horizon timeHorizon timeHorisonObst

neighborDist

 7.8% 0 þ 7.8%

 7.8% 0 þ7.8%

 7.8% 0 þ 7.8%

 18.9% 0 þ 28.9%

 18.9% 0 þ28.9%

0 0 0 0 0 0

planning scope maxNeighbors

prefVelocity

radius

 7.8% 0 þ 7.8%

0 0 0

0 0 0

 18.9% 0 þ 28.9%

 18.9% 0 þ 28.9%

0 0 0

 20.0% 0 þ 10.0%

0 0 0

0 0 0

0 0 0

 51.6% 0 þ38.7%

0 0 0

0 0 0

0 0 0

0 0 0

0 0 0

 40.0% 0 þ 60.0%

Table 4 Simulation parameters for a thin, weak and old-aged male.

Parameter values Relative Actual

timeHorizon

planning horizon timeHorisonObst

þ 21.1% 12.1 s

þ 21.1% 12.1 s

neighborDist þ 21.1% 18.2 m

planning scope maxNeighbors þ 21.1% 12

prefVelocity

radius

 51.6% 0.70 m/s

 30% 0.7 m

Fig. 9. Trajectories of persons that are all with default parameters except only one parameter. From left to right they are default, male, female, child, old-aged, thin, fat, strong and weak persons.

5. Simulation results and validation The simulation is based on Unity 3D platform, and uses RVO version 2.0. All experiments are done on a PC powered by a 3.4 GHz Intel i7-2600 CPU, a 8 GB memory and an NVIDIA GTX 560 Ti display card, and are executed in real time. Fig. 9 demonstrates walking trajectories of certain persons, only altering one of the physiological characteristics. The tiny but subtle

changes are important for agents to exhibit various behaviors in complex scenarios. Fig. 10 are obstacle avoidance scenarios, and all agents are default except four persons at the centers of the four groups. They are assigned as a fat, weak, middle-aged male at top-left subgroup, a thin, weak, old-aged, male at top-right subgroup, a fat, strong, middle-aged, female at bottom-left subgroup, and a thin, strong, child, female at bottom-right subgroup, respectively. Here we see

Please cite this article as: L. Zheng, et al., Simulating heterogeneous crowds from a physiological perspective, Neurocomputing (2015), http://dx.doi.org/10.1016/j.neucom.2014.12.103i

L. Zheng et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

7

Fig. 10. Simulation results. The top figures are agents all with default parameters, while the bottom ones are the same except four agents are different, whose trajectories are visualized.

Fig. 11. Comparison of homogeneous and heterogeneous crowds grouped by 80 individuals. The scenario is a bottleneck in front a narrow door.

the small changes of the parameters of only four agents can bring fully different results not only to motion behaviors of themselves, but also to the whole crowd. Fig. 11 demonstrates comparison between a homogeneous and a heterogeneous crowd. The individuals of the first group are all with default parameters, and those of the second group are with various simulation parameters obtained from randomly generated physiological characteristics. The results show the behaviors of individuals of the second crowd are obviously heterogeneous and reasonable. Our purpose is to simulate heterogeneous crowd behaviors on the aspect of physiological characteristics. In order to validate the proposed method, we directly compare simulated results to captured video scenes. Fig. 12 (top row) shows the real video frames at a foot bridge. Firstly, we observe carefully each

pedestrian in the video and determine their physiological characteristics including gender, age, health and body shape, as well as the appearing and vanishing positions. Then according to the observed results, mapping process is performed manually and a virtual crowd is created with heterogeneous agents owning different simulation parameters. Finally, crowd simulation is executed by setting the starting and target points of each agent based on appearing and vanishing locations. The results are depicted in Fig. 12 (middle row) with the same viewpoint as video. We also draw trajectories of three typical persons in Fig. 12 (bottom row) and find that each path is somewhat similar to real scene. Another real scene is people crossing a pedestrian street of Shibuya, Tokyo, depicted in the top row of Fig. 13. We assign pedestrians with randomly generated physiological characteristics,

Please cite this article as: L. Zheng, et al., Simulating heterogeneous crowds from a physiological perspective, Neurocomputing (2015), http://dx.doi.org/10.1016/j.neucom.2014.12.103i

L. Zheng et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

8

Fig. 12. Comparison between real scene (top row) and simulation results (middle row) is demonstrated, and trajectories of three persons are visualized (bottom row).

Fig. 13. Comparison with a crowd in the street. The top four figures are video snapshots and the bottom ones are results generated by the proposed method.

and set their goals roughly based on the video. The simulation results are depicted in the bottom row of Fig. 13. On the whole, it looks pretty similar to the real video. What is more, heterogeneous behavior of agents seems quite visually reasonable.

Acknowledgements This work was partly supported by National Natural Science Funds of China (Grant nos. 61300118, 61370167 and 61305093).

References 6. Conclusions and limitations We present a mapping model between RVO library and human physiological characteristics to simulate a virtual crowd from a physiological perspective. Mainly based on video based method, a simple, light-weight but efficient mapping method, at both single factor and combination level, is proposed to implement various motion planning based on RVO library. This paper takes into account only four physiological characteristics of human, without considering physical power, height, eyesight, etc. The mapping from physiological features to RVO parameters is not comprehensive. Also, parameter planning scope and radius calculation are simplified. All these may lead to flaws of completeness. In the future, we will continue to bring more video tracking methods to strengthen the proposed algorithm. What is more, both psychological and physiological characteristics will be combined together to simulate a real person more thoroughly.

[1] S.J. Guy, S. Kim, M.C. Lin, D. Manocha, Simulating heterogeneous crowd behaviors using personality trait theory, in: Proceedings of the 2011 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, ACM, 2011, pp. 43–52. [2] J. Van Den Berg, S.J. Guy, M. Lin, D. Manocha, Reciprocal n-body collision avoidance, in: Robotics Research, Springer, 2011, pp. 3–19. [3] L. Zheng, L. Wang, L. Liu, X. Liu, Heterogeneous crowd behaviors simulation: a physiological perspective, in: Proceedings of the Fifth International Conference on Internet Multimedia Computing and Service, ICIMCS'13, ACM, 2013, pp. 195–198. [4] C.W. Reynolds, Flocks, herds and schools: a distributed behavioral model, ACM SIGGRAPH Comput. Graph. 21 (4) (1987) 25–34. [5] D. Helbing, I. Farkas, T. Vicsek, Simulating dynamical features of escape panic, Nature 407 (6803) (2000) 487–490. [6] R. Narain, A. Golas, S. Curtis, M.C. Lin, Aggregate dynamics for dense crowd simulation, in: ACM Transactions on Graphics (TOG), vol. 28, 2009, p. 122. [7] Y. Li, M. Christie, O. Siret, R. Kulpa, J. Pettré, Cloning crowd motions, in: Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation, Eurographics Association, 2012, pp. 201–210. [8] R. McDonnell, M. Larkin, S. Dobbyn, S. Collins, C. O'Sullivan, Clone attack! Perception of crowd variety, in: ACM Transactions on Graphics (TOG), vol. 27, ACM, 2008, p. 26.

Please cite this article as: L. Zheng, et al., Simulating heterogeneous crowds from a physiological perspective, Neurocomputing (2015), http://dx.doi.org/10.1016/j.neucom.2014.12.103i

L. Zheng et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ [9] R. McDonnell, M. Larkin, B. Hernández, I. Rudomin, C. O'Sullivan, Eye-catching crowds: saliency based selective variation, in: ACM Transactions on Graphics (TOG), vol. 28, ACM, 2009, p. 55. [10] N. Hasler, C. Stoll, M. Sunkel, B. Rosenhahn, H.-P. Seidel, A statistical model of human pose and body shape, in: Computer Graphics Forum, vol. 28, Wiley Online Library, 2009, pp. 337–346. [11] W. Li, Z. Di, J.M. Allbeck, Crowd distribution and location preference, Comput. Anim. Virtual Worlds 23 (3–4) (2012) 343–351. [12] S. Takahashi, K. Yoshida, T. Kwon, K.H. Lee, J. Lee, S.Y. Shin, Spectral-based group formation control, Comput. Graph. Forum (EUROGRAPHICS 2009) 28 (2) (2009) 639–648. [13] F. Jentsch, D. Kaup, L. Malone, H. Blasko-Drabik, R. Oleson, Inclusion of social and behavioral individual-difference variables in crowd simulations: a literature review and theoretical framework, in: Proceedings of the 2008 Summer Computer Simulation Conference, Scotland, 2008, pp. 286–291. [14] F. Durupinar, N. Pelechano, J.M. Allbeck, U. Güdükbay, N. Badler, How the ocean personality model affects the perception of crowds, IEEE Comput. Graph. Appl. 31 (3) (2011) 22–31. [15] F. Durupinar, J. Allbeck, N. Pelechano, N. Badler, Creating crowd variation with the ocean personality model, in: Proceedings of the 7th International Joint Conference on Autonomous Agents and Multiagent Systems, vol. 3, International Foundation for Autonomous Agents and Multiagent Systems, 2008, pp. 1217–1220. [16] S. Curtis, S. Guy, B. Zafar, D. Manocha, Virtual tawaf: a velocity-space-based solution for simulating heterogeneous behavior in dense crowds, in: S. Ali, K. Nishino, D. Manocha, M. Shah (Eds.), Modeling, Simulation and Visual Analysis of Crowds, The International Series in Video Computing, vol. 11, Springer New York, 2013, pp. 181–209. [17] E. Ju, M.G. Choi, M. Park, J. Lee, K.H. Lee, S. Takahashi, Morphable crowds, in: ACM Transactions on Graphics (TOG), vol. 29, ACM, 2010, p. 140. [18] Q. Gu, Z. Deng, Context-aware motion diversification for crowd simulation, IEEE Comput. Graph. Appl. 31 (5) (2011) 54–65. [19] D. Kaup, T.L. Clarke, R. Oleson, L. Malone, F.G. Jentsch, Introducing age-based parameters into simulations of crowd dynamics, in: Simulation Conference, WSC 2008, Winter, IEEE, 2008, pp. 895–902. [20] N. Pelechano, J.M. Allbeck, N.I. Badler, Controlling individual agents in highdensity crowd simulation, in: Proceedings of the 2007 ACM SIGGRAPH/ Eurographics Symposium on Computer Animation, Eurographics Association, 2007, pp. 99–108. [21] M.G. Conner, Understanding the Difference Between Men and Women, 〈http:// www.oregoncounseling.org/articlespapers/documents/differencesmen women.htm〉. [22] R. Lobjois, V. Cavallo, The effects of aging on street-crossing behavior: from estimation to actual crossing, Accid. Anal. Prevent. 41 (2) (2009) 259–267. [23] J.L. Purser, M. Weinberger, H.J. Cohen, C.F. Pieper, M.C. Morey, T. Li, G.R. Williams, P. Lapuerta, Walking speed predicts health status and hospital costs for frail elderly male veterans, J. Rehabil. Res. Dev. 42 (4) (2005) 535–546. [24] R. Hong, M. Wang, Y. Gao, D. Tao, X. Li, X. Wu, Image annotation by multipleinstance learning with discriminative feature mapping and selection, IEEE Trans. Cybern. 44 (5) (2014) 669–680. [25] Z.-J. Zha, L. Yang, T. Mei, M. Wang, Z. Wang, T.-S. Chua, X.-S. Hua, Visual query suggestion: towards capturing user intent in internet image search, ACM Trans. Multimed. Comput. Commun., Appl. (TOMCCAP) 6 (3) (2010) 13. [26] K. Zhang, L. Zhang, M.-H. Yang, Real-time compressive tracking, in: Computer Vision, ECCV 2012, Springer, 2012, pp. 864–877. [27] M. Wang, R. Hong, X.-T. Yuan, S. Yan, T.-S. Chua, Movie2comics: towards a lively video content presentation, IEEE Trans. Multimed. 14 (3) (2012) 858–870. [28] Z.-J. Zha, M. Wang, Y.-T. Zheng, Y. Yang, R. Hong, T.-S. Chua, Interactive video indexing with statistical active learning, IEEE Trans. Multimed. 14 (1) (2012) 17–27. [29] Z.-J. Zha, H. Zhang, M. Wang, H. Luan, T.-S. Chua, Detecting group activities with multi-camera context, IEEE Trans. Circuits Syst. Video Technol. 23 (5) (2013) 856–869.

9 Dang Qin is a master student in school of computer and information, Hefei University of Technology. His research interests include crowd simulation and visualization.

Yajun Cheng is a master student in school of computer and information, Hefei University of Technology. His research interests include crowd simulation and formation animation.

Lin Wang is a master student in school of computer and information, Hefei University of Technology. His research interests include virtual reality and simulation.

Lin Li is a lecturer in school of computer and information, Hefei University of Technology. Her research interests include visualization and virtual reality.

Liping Zheng is an associate professor in school of computer and information, Hefei University of Technology. His research interests include visualization and computer simulation.

Please cite this article as: L. Zheng, et al., Simulating heterogeneous crowds from a physiological perspective, Neurocomputing (2015), http://dx.doi.org/10.1016/j.neucom.2014.12.103i