Machine Learning for Improved Crowd Simulations (Advert Reference: SF22/EE/CIS/AMOS)

Crowd simulations now play an integral role in performance-based fire safety design, and they are applied in a wide variety of domains, from evacuation planning and management to incident response and analysis Although a number of commercial and open-source software tools exist to simulate crowd movement and behaviour, the underlying “engine” (which determines how simulated people move through space and interact with one another and their environment) may still generate "unrealistic" crowd behaviours. This “reality gap” presents a significant challenge in terms of the adoption of policies based on simulation visualisations; put simply, decision makers (event organizers, local authorities, architects, etc.) may not entirely trust the outputs of these models.

In this project, we build on our recent work in this area [2,3] and propose three main work components: (1) the use of machine learning techniques to automatically detect signature features of real crowds, and (2) the application of these techniques to much larger and more realistic datasets (eg. the Stanford Drone dataset [4], which includes many more types of movement, spatial diversity, etc.) This facilitates (3) the integration of identified features into an open-source crowd simulation package, in order to establish whether or not their inclusion leads to the generation of more realistic simulated crowds. Taken together, this work will lead to a significant evolution of crowd simulation, with potential high impact in terms of policy, urban design, and human safety.

This work may be conducted in collaboration with a number of existing industrial partners (eg. pedestrian movement consultants and providers of crowd simulation software).

For informal enquiries please contact Martyn Amos ([Email Address Removed])

[1] Boje, C., & Li, H. (2018). Crowd simulation-based knowledge mining supporting building evacuation design. Advanced Engineering Informatics, 37, 103-118.

[2] Webster, J. and Amos, M. (2020). A Turing test for crowds. Royal Society Open Science 7:200307. doi: 10.1098/rsos.200307. 

[3] Webster, J. and Amos, M. (2021). Identification of lifelike characteristics of human crowds through a classification task. Proc. Conference on Artificial Life (ALIFE2021), Prague, Czech Republic, Jul. 19-23 2021. doi:10.1162/isal_a_00360.

[4] Stanford Drone Dataset, https://cvgl.stanford.edu/projects/uav_data/

Eligibility and How to Apply: 

Please note eligibility requirement:  

·               Academic excellence of the proposed student i.e. 2:1 (or equivalent GPA from non-UK universities [preference for 1st class honours]); or a Masters (preference for Merit or above); or APEL evidence of substantial practitioner achievement. 

·               Appropriate IELTS score, if required. 

For further details of how to apply, entry requirements and the application form, see 

https://www.northumbria.ac.uk/research/postgraduate-research-degrees/how-to-apply/  

Please note: Applications that do not include a research proposal of approximately 1,000 words (not a copy of the advert), or that do not include the advert reference (e.g. SF22/…) will not be considered. 

Start Date: 1 October 2022 

Northumbria University takes pride in, and values, the quality and diversity of our staff and students. We welcome applications from all members of the community.