Introduction: The safety of humans in crowded environments has been recognised as an important and rapidly growing research area with significant implications for urban planning, event management, building design, fire safety engineering and rescue service to name a few. This stream of research is aimed at guiding safe designs and effective evacuation plans by simulating emergency scenarios and estimating measures such as total evacuation time. A large body of research has also been dedicated to the development of modelling tools with the capability to identify (and thus prevent) circumstances that lead to crowd discomfort, crashes or disasters in mass gatherings and public facilities. It has, however, been argued that the empirical knowledge in this area has lagged behind the theoretical developments and computational capabilities. This has left the descriptive power of the existing models for reproducing the natural behaviour of humans questionable given that in many cases there is a lack of reliable and well-conditioned data for model validation or calibration purposes. Methods: With the vast majority of the empirical knowledge in this fast-growing and interdisciplinary field being very recent, a survey of the existing literature is still missing. Here, we gather together the existing empirical knowledge in this area in a comprehensive review (based on surveying more than 160 studies restricted to those published in peer-reviewed journals since 1995) in order to help bridge this gap. We introduce for the first time a categorisation system of the relevant data collection techniques by recognising seven general empirical approaches. We also differentiate between various aspects of human behaviour pertinent to crowd behaviour by putting them into perspective in terms of three general levels of "decision making". We also discuss the advantages and disadvantages offered by each data collection technique. Major gaps and poorly-explored topics in the current literature are discussed. Findings and applications: Our major conclusion is that the empirical evidence in this area is largely disperse and even in some cases mixed and contradictory, requiring a more unified system of terminologies and problem definitions as well as unified measurement methods in order for the findings of different studies to become replicable and comparable. We also showed that the existing body of empirical studies display a clear imbalance in addressing various aspects of human behaviour with certain (but crucial) aspects (such as "pre-movement time" and "choice of activity") being poorly understood (as opposed to our knowledge and amount of data about "walking behaviour" for example). Our review also revealed that previous studies have predominantly displayed a stronger tendency to study the behaviour based on aggregate measures as opposed to individual-level data collection attempts. We hope that this collection of findings sets clearer avenues for advancing the knowledge in this area, guides future experiment designs and helps researchers form better-informed hypotheses and choose most suitable data collection methods for their question in hand. (C) 2017 Elsevier Ltd. All rights reserved.
