Assessing the perceived realism of agent grouping dynamics for adaptation and simulation

Highlights

• The average absolute thresholds for group frequency are 15 groups and 36 groups.

• The average absolute thresholds for group density are 1.6 agents per group and 3.8 agents per group.

• The average optimum configuration for group frequency is 24 groups.

• The average optimum configuration for group density is 2.3 agents per group.

• There is no change in participant’s responses when measured at the three locations.

Abstract

Virtual crowds are a prominent feature for a range of applications; from simulations for cultural heritage, to interactive elements in video games. A body of existing research seeks to develop and improve algorithms for crowd simulation, typically with a goal of achieving more realistic behaviours. For applications targeting human interaction however, what is judged as realistic crowd behaviour can be subjective, leading to situations where actual crowd data is not always perceived to be more real than simulation, making it difficult to identify a ground truth. We present a novel method using psychophysics to assess the perceived realism of behavioural features with respect to virtual crowds. In this instance, a focus is given to the grouping dynamics feature, whereby crowd composition in terms of group frequency and density is evaluated through thirty-six conditions based on crowd data captured from three pedestrianised real-world locations. The study, conducted with seventy-eight healthy participants, allowed for the calculation of perceptual thresholds, with configurations identified that appear most real to human viewers. The majority of these configurations correlate with the values extracted from the crowd data, with results suggesting that viewers have more perceptual flexibility when group frequency and density are increased, rather than decreased.

Introduction

Realistic crowd simulation is complex and a broad topic for research, due in part to the many elements that comprise a system for producing emergent crowd behaviour [38], [66]. Similarly contributing, is the fact that virtual crowds are utilised for a wide range of applications [4]. We make a distinction between serious applications and those for entertainment, as both are trying to achieve different end results and thus typically utilise different design approaches. Examples of serious applications include simulations for evacuation procedures [1], virtual crowds for specific scenarios such as bioterrorism [63], and simulations for the reconstruction of cultural heritage [67]. These types of applications are often modelled using crowd data [68] and have a limited amount of human interaction, being built towards specialised purposes. In contrast, virtual crowds utilised for entertainment are typically geared towards a user-centered focus, wherein user interaction and perception are highly important considerations [52]. A clear example can be seen looking at crowds applied as interactive elements in video games, with developers stressing the importance of believability or perceived realism [10]. Serious games on the other hand, offer a middle ground between a serious focus and entertainment, often borrowing elements from both in their design approach. This differentiation of purpose is crucial, both for research and implementation, especially for crowd simulations designed for user experience where the development process is often highly constrained in terms of time and computational resources [15]. The choices developers make greatly impact the resultant crowd behaviour and can be the difference between achieving plausible crowds or not.

In this context, crowd behaviour can be defined as the overall outwardly visible results of the underlyWhen viewing a crowd, some of the most noticeable elements are directly related to the group dynamics. “How many individual groups make up the crowd?” “How many people are in each of these groups?” These are important questions when analysing a crowd at large and are an equally important consideration when implementing a virtual crowd. The number of groups and their density has the potential to alter how the crowd is perceived. In some urban locations, a crowd consisting of a lot of highly dense groups can considered more perceptually realistic than a lot of low density groups for example. In others a few highly-dense groups might be considered more perceptually realistic than a lot of highly dense groups. Both frequency and density of groups have the potential to alter the overall composition of the virtual crowd and thus change how it is perceived, building upon previous studies that highlight crowd composition as an important factor [55]. As such there is room for the exploration of these two key variables to highlight certain perceptual thresholds regarding which values are considered perceptually realistic.ing artificial intelligence algorithms employed in a simulations design [17], [2]. This emergent crowd behaviour thus incorporates the combined outcome of agent internal motivations and reactions between crowd members, as well as the virtual environment, and other potential stimuli based on the configuration. This crowd behaviour can be broken down further into the concept of behavioural features, whereby specific elements of the resultant behaviour can be identified as focal points [50].

Typically, a crowd system will include a form of decision making [41], pathfinding [27], [16] and steering [58], which allow agents to perceive, think, and act, to a limited extent. Sometimes, however, these base algorithms do not produce crowd behaviour realistic enough for a simulation to realise its purpose [59]. In these instances, additional algorithms can be implemented to impose different effects upon the resultant crowd behaviour [30], [8], [17]. For example, by implementing a Social Forces Model [29] that seeks to replicate social forces between agents, certain dynamics are created. Agents can be attracted or repelled from one another depending upon the specific forces modelled and the weight factors involved, all of which results in different crowd behaviour. As such, the choices made when implementing a crowd system greatly determines the resultant crowd behaviour and particularly in cases involving user interation, it can be difficult to know how the results will be perceived [18], [12].

There is a body of existing research for developing and improving artificial intelligence algorithms for crowd based simulation [42], [6], [36]. Some research aims to utilise new techniques or models to simulate crowd behaviour [9], [28], [33], whilst other works extend existing behaviours to make them more realistic [37], [47]. These lines of research typically result in new and interesting behaviours or refinements, but often do not consider effectiveness from a user perspective [69]. Industry developers of AI for commercial purposes such as video games, particularly note that “AI is always a minefield [25, p.154], with various examples of ineffective agents [25]. A recent project, for example, that has explored the value of authentic crowd simulation AI in a cultural heritage context, was the Inside Joycean Dublin Project and later the I-Ulysses: Poetry in Motion project, with which the project is associated [14].

Realism in terms of simulation and virtual environments is an important metric [40], [34], [35], [7], particularly with efforts in research and development seeking to produce realistic virtual crowds [31], [23]. Additionally, realism has been shown to an influential factor in a number of arenas, including immersion, engagement, and education [65], [11]. As noted however, depending on the application of the virtual crowd, the requirements in terms of realism can vary [26]. For serious applications, simulated behaviour needs to be as close to that of a real crowd as possible; this ensures results are accurate for reasoning about real-world environments, for example in areas such as construction and safety [27], [3], [49]. Crowd behaviour exhibited by an application for entertainment however, needs to be perceptually plausible to the users to aid interaction and immersion [46], [13], which is not necessarily the case for current high-fidelity reproductions of real-world behaviours. In these cases, realism is subjective and influenced by perception, rather than being true to what is actually realistic [62]. It has been shown that simulation correctness based on actual crowd data is not guaranteed to be the most perceptually realistic configuration for human viewers and that it is possible for parametrised and synthesised behaviour to produce equal or better results [55].

In the case of virtual crowds intended for user interaction, since neither simulation correctness based on crowd data, or algorithmic complexity, are universally applicable as metrics for perceived realism, perceptual evaluation can instead be utilised to assess the effectiveness of simulated crowd behaviour. Such methods are more generally used to evaluate graphical applications, and research exists showing the benefit provided by probing human perception [21], [44], [56]. The use of psychology, in particular, is a fundamental component under paradigms whereby the end-goal of a visualisation is the human response, rather than a rendered image. One such psychological method is psychophysics [5].

“Psychophysics is commonly defined as the quantitative branch of the study of perception, examining the relations between observed stimuli and responses and the reasons for those relations [5, p.1]. This method has the key benefit of allowing for the calculation of perceptual thresholds through examination of the psychometric function. These thresholds being fundamental to psychophysics [22] are the reason for its application in auditory and visual stimuli identification experiments [48]. However, its application to virtual crowds, and in particular crowd behaviour, has been limited to non-existent, with most studies applying more traditional perceptual evaluation techniques [55]. However, by combining the use of crowd data with psychophysics through what we refer to here as comparative psychophysics, it adds a potential new dimension to assess and refine resultant crowd behaviour. This comes from identification of the absolute thresholds and being able to compare human perception to the reality of the situation through the use of this crowd data. As we have noted, the choices made when implementing a virtual crowd greatly impact the resultant crowd behaviour and thus it is of benefit to inform developers through these ‘perceptual metrics, particularly when considering user-centered applications.

Realism is shown to be an important aspect of crowd behaviour and different applications necessitate differing qualities and definitions of realism. It is clear that the efficacy of simulated crowd behaviour for user-centered applications must be evaluated against something more sophisticated than just adherence to crowd data and simulation correctness. This work is a proof of concept for the novel methodological approach of applying aspects of psychophysics for threshold calculation towards elements of simulated crowd behaviour. To this end, we examine grouping dynamics, consisting of two variables, group frequency and group density. This selection is due to a number of factors, including the features prominence and wide-applicability to most all crowd simulation scenarios. Thus, given this objective and our focus on grouping dynamics for crowd simulation, we present four research questions:

• RQ1: What are the thresholds for group frequency and group density to be considered perceptually plausible?

• RQ2: What composition of group frequency and group density provides the highest levels of perceived realism?

• RQ3: What effect does location have on the perceived realism of group frequency and group density?

• RQ4: What are the differences for crowd compositions, if any, between crowd data and perceived realism?

By considering these four questions, an outcome of this work is to also provide perceptual metrics in the form of thresholds and optimum configurations to inform and support crowd simulation development, in which human interaction and acceptance is of intrinsic value.

Towards this goal, in this paper, we present here a novel framework for employing a comparative psychophysical method to evaluate the perceived realism of behavioural features. Through a three step-methodology, the grouping dynamic feature is identified, implemented into the urban crowd simulation (Fig. 1) and and subjected to a quantitative psychophysical experiment using crowd data from three real-world locations. The perceptual thresholds are calculated, the optimum configurations examined and the differences between crowd data and the analysed perceptions are explored. The method and novel application of comparative psychophysics towards behavioural features, is to inform algorithm choice and allow for development flexibility while remaining within the bounds of plausibility through threshold identification, ensuring user requirements are kept the forefront but giving implementation options when resources are constrained.

The remainder of this paper is structured as follows: Section 2 considers related work. Section 3 highlights the methodology and framework. The analysis of crowd footage is detailed in Section 4. Section 5 covers implementation of the crowd simulation through synthesis. In Section 6 perception is outlined with the psychophysics experiment and its results. Finally, Section 7 provides conclusions and areas for further research.