New learning based method to model crowd behaviour in simulated environments

A team of scientists from CYENS Centre of Excellence and the University of Cyprus, led by Dr Panayiotis Charalambous have created a new learning based method to model crowd behaviour in simulated environments.

The proposed solution allows to model the diversity among agents’ behaviors and achieve heterogeneity in virtual crowds. To achieve that, a Reinforcement Learning-based framework has been developed to simulate multiple agents’ behaviors concurrently: heterogeneous behaviors and their various combinations and interactions were considered to form diverse agent profiles within a virtual crowd. The agents’ behavior can be optimised by cleverly varying the importance of several common behavior types of sub-behaviors, namely goal seeking, collision avoidance, interactions with points of interest in an environment, and grouping while training the model. The importance of each separate sub-behavior is added as input to the agent policy, resulting in the development of a unified crowd simulation model capable of both capturing and enabling dynamic run-time manipulation of several different agent profiles. Thus, this method allows creating fully configurable profiles within a virtual crowd to ensure realism and plausibility in the resulting simulated scenarios.

Through a series of experiments, the authors verify that the system provides users with the ability to design realistic virtual scenes, control and mix agent behaviors as well as creating personality profiles and assign different profiles to groups of agents. Interestingly, this new learning model can handle situations not seen in the training data such as crowds with higher density, behavior combinations that are outside the training intervals and scenes with more intricate environment layouts.

This work has been accepted at the ACM SIGGRAPH Conference, the highest-tier international platform for disseminating cutting edge scholarly work in computer graphics technology and interactive techniques. This year, SIGGRAPH will be organized in Vancouver, Canada. The scientific excellence of the ideas is the predominant acceptance criterion. The SIGGRAPH conference proceedings has the highest impact factor among academic publications in the field of computer graphics. With an acceptance rate ranging between avaraging 20%, only 200 papers from all over the world in the general area of computer graphics, animation, and visualization achieve this status. SIGGRAPH has an average attendance of over 15000 people with more than 150 industry exhibitors; this makes it one of the largest technical conferences in the world.