Crowd Behavior Analysis: A Review where Physics meets Biology

1. Introduction

• Crowds and swarms can exhibit emergent collective behavior through simple interactions, while human crowds may spontaneously form coordinated walking lanes without communication.The review notes that physics- and biology-inspired approaches share perspectives but are not clearly distinguished.Source: §1 · pp. 1–2

1.1. Comparisons with Previous Reviews

• Prior surveys emphasize computer-vision techniques or physics-based models, while the authors identify no comprehensive review of biologically inspired computer-vision algorithms for crowd analysis.The authors note that many methods already apply biological concepts.Source: §1.1 · pp. 2–3

1.2. Motivation and Contributions

• The review synthesizes state-of-the-art computer-vision methods for crowd behavior analysis through physics- and biologically inspired perspectives, covering crowd segmentation, dynamic analysis, and density estimation.It first examines common and conflicting attributes, then discusses applications and benchmarks, aiming to encourage multidisciplinary integration.Source: §1.2 · p. 6

2.1. What is a swarm or crowd?

• Swarm commonly refers to spatially proximate entities, with biology associating swarms with social insects and physics using the term for particles or electrons in transport equations and fluid flows.Biology-inspired examples include Ant Colony Optimization and Particle Swarm Optimization.Source: §2.1 · p. 7
• Reynolds's Boids model used separation, alignment, and cohesion rules to simulate coordinated flock movement and illustrate how simple interactions can produce complex emergent behavior.The model helped spur interest in adapting emergent behavior to complex computer vision problems.Source: §2.1 · p. 7, Figure 3 · p. 7

2.2. What is the role of computer vision in crowd analysis?

• Monitoring crowds of hundreds or thousands is difficult because operators face many screens, fatigue, distractions, and complex behavior, while missed overcrowding or suspicious activity can have catastrophic consequences.Crowds may plunge into panic under physical stress such as overcrowding or sudden external pressure such as shootings or fire.Source: §2.2 · p. 8
• Carnage in crowds doubled over the past two decades, and disaster investigations identified missed opportunities for technology to support proactive crowd surveillance.The survey focuses on physics- and biology-inspired methods for crowd analysis in computer vision.Source: §2.2 · p. 8

3.1. Decentralized

• Decentralized interactions can generate collective behavior without leadership, but leader-based models and a gap between centralized/decentralized models and real-world crowds challenge this account.Decentralized mechanisms involve feedback and multiple agent interactions; evacuation models coexist with leader- or subgroup-based approaches.Source: §3.1 · p. 10

3.2. Collective motion

• Collective motion appears in living and physical systems, and crowd models combine cognitive heuristics or motivation-based social forces to represent individual movement and interaction.Collective motion is ordered macroscopic behavior, while people can act as a socially coherent mass without prior awareness.Source: §3.2 · p. 11
• Collective-motion methods detect abnormalities when individuals deviate from crowd flow; individual uniqueness affects crowd-model scalability and robustness, while local and global group and environmental interactions further complicate analysis.Source: §3.2 · p. 12

3.3. Emergent behavior

• Emergent behavior is complex pattern formation from simple rules, distinct from collective motion, and can arise without external coordination or leaders.An ant colony, despite the limited memory and simple actions of individual ants, can find the shortest route to food.Source: §3.3 · pp. 12–13
• Crowd-emergence methods classify behaviors such as bottleneck, lane, arch, fountainhead, and blocking, but predefined patterns and rules may not capture real-world complexity.Allain et al. model seven crowd behaviors using Lagrangian forces.Source: §3.3 · p. 13

4.1. Bias or non-bias

• Despite individual differences, crowds may converge on a majority decision, and people adjust judgments to align with crowd opinions.Individuals share a focal point but have varying self-interest and inclinations; the passage links conformity to social norms with reinforcement learning and neuronal responses.Source: §4.1 · pp. 13–14
• Models can relax individual bias to emphasize collective effects; approaches use crowd patterns or conformity to infer individual behavior, but real-world conformity's extent remains open.The passage identifies the challenge of modeling each entity's varying predilections and behavior as complex.Source: §4.1 · pp. 13–15

4.2. Thinking or non-thinking

• Physics-based models treated crowds as homogeneous, non-thinking particles moved by external forces, but omitted individual reactions can make crowd-flow models imprecise and hazardous.A person who falls may obstruct flow and contribute to stampede and deaths; newer approaches combine physical and psychological aspects at both crowd and individual scales.Source: §4.2 · p. 15
• Because computational complexity grows with crowd density, models must balance individuals' decision-making capacity against the influence their behavior has on others.The passage calls for further investigation of thinking and non-thinking, and bias and non-bias, to better mitigate harmful emergent behavior.Source: §4.2 · p. 16, Figure 12 · p. 16
• In smoke-room evacuation, Helbing et al. [32] modeled herding and exit clogging where fluid-flow models predict efficient use of both exits, by combining individualistic and collective behavior.This result is presented as closer to actual evacuation behavior and as preliminary evidence motivating further study of thinking, non-thinking, bias, and non-bias in crowds.Source: Figure 12 · p. 16, §4.2 · p. 16

5.1. Dataset

• Because research approaches and benchmarks developed independently across disciplines, evaluation protocols and performance comparisons are difficult to summarize, motivating a common platform for crowd analysis.The section also catalogs publicly available datasets for crowd behavior analysis.Source: §5.1 · p. 17

5.2. Crowd segmentation

• Crowd segmentation groups coherent motion, often using whole-scene flow fields, while trajectory tracking and mesoscopic models capture individual paths or local interactions.Tracking is difficult in crowds because interactions, occlusions, and background clutter fragment trajectories; segmentation also supports anomaly detection by identifying deviations from regular flow.Source: §5.2 · p. 20

5.3. Crowd dynamic analysis

• High motion dynamics and irregularities can indicate salient behavior, while dense crowds following dominant flows tend to have stable dynamics.Approaches include modeling local interactions or group properties and learning normal behavior, but surveillance analysis is constrained by the scarcity of abnormal-event training data; some non-learning methods target specific events.Source: §5.3 · pp. 21–22

5.4. Crowd density estimation

• Crowd density estimation uses collective-motion and appearance cues rather than individual intent, but increasing density and occlusion make people harder to distinguish and can cause trajectory clusters to merge.Appearance-based methods avoid detecting every person, while regression approaches face perspective distortion; segmentation and head-shoulder detection can also be vulnerable to inter-occlusion.Source: §5.4 · pp. 23–24

Stationary crowd

• Stationary-crowd analysis is underexplored despite non-motion cues revealing groups that can obstruct and alter scene dynamics; methods detect gathering, stopping, relocating, and deforming.An earlier system detects abandoned objects despite occlusion when the object is visible for 50% of the time.Source: §6 · p. 25

Still images

• Image-based analysis can complement video’s temporal information by separating crowd-appearance cues from background clutter, and can address stationary or dynamic crowds.Most crowd-behavior studies focus on video, while only limited areas use single images.Source: §6 · pp. 25–26

Large-scale

• The review anticipates deep learning for crowd behavior analysis, drawing on large-scale CCTV data while noting a shift from black-box models toward high-dimensional data transforms.Deep learning can learn representations automatically from large quantities of data, but its application to crowd behavior remains an emerging opportunity.Source: §6 · p. 26
• Deep crowd models remain domain-dependent, with motion modeling, transfer learning, and further technical investigation identified as unresolved opportunities [174].In [174], motion filters are pre-trained before joint optimization with appearance filters; other attempts address crowd-scene understanding and density estimation.Source: §6 · p. 26

• The survey avoids ranking physics over biology or vice versa, noting their resemblance and distinction remain vague amid complex human behavior while observing a paradigm shift toward integration.Source: §7 · p. 26
• The survey finds physics and biologically inspired approaches complementary, but their fusion for crowd-behavior analysis remains underused, calling for cross-disciplinary bridging to build “living” computer vision systems.Source: §7 · p. 26
• Human behavior may exceed what these two approaches alone can model, motivating multidisciplinary work and future research that bridges current solutions with real-world requirements.The survey identifies statistical physics, sociology, philosophy, and other life sciences as potential contributors, and describes the field as emerging and worthy of further investigation.Source: §7 · p. 27