Search for: All records

Abstract Global patterns of collective motion in bird flocks, fish schools, and human crowds are thought to emerge from local interactions within a neighborhood of interaction, the zone in which an individual is influenced by their neighbors. Both metric and topological neighborhoods have been reported in animal groups, but this question has not been addressed for human crowds. The answer has important implications for modeling crowd behavior and predicting crowd disasters such as jams, crushes, and stampedes. In a metric neighborhood, an individual is influenced by all neighbors within a fixed radius, whereas in a topological neighborhood, an individual is influenced by a fixed number of nearest neighbors, regardless of their physical distance. A recently proposed alternative is a visual neighborhood, in which an individual is influenced by the optical motions of all visible neighbors. We test these hypotheses experimentally by asking participants to walk in real and virtual crowds and manipulating the crowd's density. Our results rule out a topological neighborhood, are approximated by a metric neighborhood, but are best explained by a visual neighborhood that has elements of both. We conclude that the neighborhood of interaction in human crowds follows naturally from the laws of optics and suggest that previously observed “topological” and “metric” interactions might be a consequence of the visual neighborhood. 

Where does the organization in behavior come from? Ecological psychology seeks to explain adaptive behavior as self-organized, emerging from the dynamics of agent-environment interactions, without assuming such organization a priori. This chapter offers an accessible introduction to the behavioral dynamics approach to modeling perception and action, applied to the case of human locomotion and crowd behavior. Most models of such behavior are ‘omniscient’, presuming accurate knowledge of the environmental state, but we find that information-based ‘visual’ models more closely capture human data. At the individual level, a locomotor trajectory emerges from an agent’s interactions with goals and obstacles, as attractors and repellers appear, shift, and bifurcate. At the dyad level, pedestrian interactions governed by visual control laws yield coordinated movements such as interception, collision avoidance, and following. At the collective level, these local interactions generate human ‘flocking’, crowd bifurcations, and patterns of lanes and stripes in crossing flows of pedestrians. By adopting an empirical, bottom-up, information-based approach, behavioral dynamics helps explain individual and collective behavior as resulting from a process of self-organized pattern formation, without appealing to internal plans or external commands. 

Road tunnels are enclosed spaces that most occupants only experience while driving through them. In case of fire, however, occupants potentially need to evacuate on foot from a dangerous and unfamiliar environment. Clear and accurate guidance is important for an efficient and safe evacuation from tunnels. Common cues for evacuation guidance are a signage and audio messages that attract occupants to move on appropriate egress routes and avoid unsafe routes. This paper investigates how different types of visual and auditory signals influence occupants’ exit choices in a simulated tunnel evacuation. Common guidance cues were presented to participants in a mobile Head Mounted Display, and they were asked to choose between two possible exit doors in a simulated road tunnel. Two attracting cues (‘‘EXIT’’ signs, audio instructions), and two detracting cues (‘‘DO NOT ENTER’’ signs; traffic cones placed in front of an exit) were studied in three virtual reality (VR) experiments. In each experiment, the presence and direction of the cues were manipulated, and data from 20 participants were collected. Experiment 1 explored the effects of attracting cues, Experiment 2 detracting cues, and Experiment 3 the combination of attracting and detracting cues. Across all studies, participants tended to follow the guidance provided when there was only one cue. When several competing and even contradictory cues were present, participants were most likely to rely on audio instructions, followed by traffic cones and ‘‘DO NOT ENTER’’ signs, whereas ‘‘EXIT’’ signs were often disregarded. We conclude that participants tend to follow temporary cues that could carry current information, as opposed to permanently installed signage. Some corresponding suggestions are put forward on evacuation system design and strategic planning in a tunnel fire. 

Previous simulations of crossing flows using a vision-based collision-avoidance model reproduced lanes and stripes but showed larger heading adjustments during crossing than the human data. Here we investigate two possible explanations. First, we tested participants walking through a virtual crowd under two density conditions, refit the collision avoidance model, and re-simulated the crossing flows data. Our findings reveal little influence of moderate densities on human collision avoidance behavior. Second, we are testing mutual collision avoidance between two participants to determine whether a revised model better approximates the crossing flows data. 

We previously reported an experiment in which covert or explicit leaders (confederates) were placed in a group of walking pedestrians in order to test leader influence on human crowd motion. Here we simulate the participant trajectories with variants of an empirical pedestrian model, treating the covert leaders’ motion as input, and test model agreement with the experimental data. We are currently using reconstructed influence networks [2] to modify the model weights in order to simulate the influence of explicit leaders. The results help us to understand how leader influence propagates via local interactions in real human crowds. 

The visual control of locomotion has been modeled for individual pedestrian behavior; however, this approach has not been applied to collective human behavior, where spontaneous pattern formation is often observed. We hypothesize that an empirical visual model of human locomotion will reproduce the emergent pattern of lanes and stripes observed in crossing flows, when two groups of pedestrians walk through each other at crosswalks or intersections. Mullick, et al. (2022) manipulated the crossing angle between two groups and found an invariant property: stripe orientation is perpendicular to the mean walking direction (i.e. 90˚ to the bisectrix of the crossing angle). Here we determine the combination of model components required to simulate human-like stripes: (i) steering to a goal (Fajen & Warren, 2003), (ii) collision avoidance with opponents (Bai, 2022; Veprek & Warren, VSS 2023), and (iii) alignment with neighbors (Dachner, et al., 2022), together called the SCruM (Self-organized Collective Motion) model. We performed multi-agent simulations of the data from Mullick et al. (2022), using fixed parameters and initial conditions from the dataset. There were two sets of participants (N=36, 38) with 18 or 19 per group. Crossing angle varied from 60˚ to 180˚ (30˚ intervals), with ~17 trials per condition. The minimal model necessary to reproduce stripe formation consists of the goal and collision avoidance components. Mean stripe orientation did not differ from 90˚ to the bisectrix (BF10 < 0.01, decisive). However, the SD of heading during crossing was significantly larger than the human data (p<0.001), whereas the SD of speed was significantly smaller (p<0.001). Thus, the ratio of heading/speed adjustments is lower than previously found, implying the need to reparameterize model components for walking in groups. In sum, steering to a goal and collision avoidance are sufficient to explain stripe formation in crossing flows, while alignment is unnecessary. 

Efficient emergency guidance in buildings is essential for the safe evacuation of occupants. However, occupants may be exposed to contradictory information from signage and other sources of information. This study presents a set of forced-choice VR experiments and a machine learning approach to investigate the effect of competing or conflicting guidance on exit choice in simulated scenarios. In the VR study, participants chose between two potential exits under time pressure in each trial. Attracting cues (“EXIT” signs, audio instructions) and repelling cues (“DO NOT ENTER” signs, traffic cones) were placed in front of the two exits, either individually or in combination. In total, 2,125 datapoints were recorded from 20 participants. To model exit choice, machine learning (random forest, RF) models were applied to predict and interpret the guidance on evacuation choices. The tuned-hyperparameters RF model proposed in this study showed above 75% accuracy to predict evacuation choices facing conflict cues and was superior to default RF and logistic regression models. Interestingly, repelling cues such as “DO NOT ENTER” signs had a stronger impact on exit choice than attracting cues like “EXIT” signs when people have to make choices. Overall, the study offers valuable data and insights into exit choices, revealing that negative cues are more influential than positive ones in emergencies. These findings can significantly inform the design and optimization of egress guidance systems. This bias towards negative information under pressure suggests that evacuation systems should prioritize clear and prominent negative cues to guide occupants effectively. 

Collective motion in human crowds has been understood as a self-organizing phenomenon that is generated from local visual interactions between neighboring pedestrians. To analyze these interactions, we introduce an approach that estimates local influences in observational data on moving human crowds and represents them as spatially-embedded dynamic networks (visual influence networks). We analyzed data from a human “swarm” experiment (N= 10, 16, 20) in which participants were instructed to walk about the tracking area while staying together as a group. We reconstructed the network every 0.5 seconds using Time-Dependent Delayed Correlation (TDDC). Using novel network measures of local and global leadership ('direct influence' and 'branching influence'), we find that both measures strongly depend on an individual’s spatial position within the group, yielding similar but distinctive leadership gradients from the front to the back. There was also a strong linear relationship between individual influence and front-back position in the crowd. The results reveal that influence is concentrated in specific positions in a crowd, a fact that could be exploited by individuals seeking to lead collective crowd motion. 

For pedestrians moving without spatial constraints, extensive research has been devoted to develop methods of density estimation. In this paper we present a new approach based on Voronoi cells, offering a means to estimate density for individuals in small, unbounded pedestrian groups. A thorough evaluation of existing methods, encompassing both Lagrangian and Eulerian approaches employed in similar contexts, reveals notable limitations. Specifically, these methods turn out to be ill-defined for realistic density estimation along a pedestrian’s trajectory, exhibiting systematic biases and fluctuations that depend on the choice of parameters. There is thus a need for a parameter-independent method to eliminate this bias. We propose a modification of the widely used Voronoi-cell based density estimate to accommodate pedestrian groups, irrespective of their size. The advantages of this modified Voronoi method are that it is an instantaneous method that requires only knowledge of the pedestrians’ positions at a give time, does not depend on the choice of parameter values, gives us a realistic estimate of density in an individual’s neighborhood, and has appropriate physical meaning for both small and large human crowds in a wide variety of situations. We conclude with general remarks about the meaning of density measurements for small groups of pedestrians. 

Crossing flows of pedestrians result in collective motions containing self-organized lanes or stripes. Over a wide range of crossing angles, stripe orientation is observed to be perpendicular to the mean walking direction. Here, we test the behavioral components needed to reproduce the lanes and stripes in human data using an empirical, vision-based pedestrian model (Visual SCruM). We examine combinations of (i) steering toward a goal, (ii) collision avoidance, and (iii) alignment (both with and without visual occlusion). The minimal model sufficient to reproduce perpendicular stripes was the combination of a common goal and collision avoidance, although the addition of alignment with occlusion better approximated human heading adjustments. However, the model overestimated the variation in heading and underestimated the variation in speed, suggesting that recalibration of the collision avoidance component is needed.