Optimizing k-NN Queries for Dense Interactive Real-time Crowd Simulations

Abstract

As crowd simulations become an ever more important aspect of managing events, city growth or pandemics, it is important that the performance of software solutions keep up. In this thesis we demonstrate a novel grid-based memory resident index structure that massively outperforms the state-of-the-art competitors. This is achieved by utilising properties intrinsic to dense crowd simulations. We’ve started with a basic implementation and improved on it by employing new techniques such as iterative enlarging and field of view while keeping the datastructure simple for high performance construction. Acceleration structure construction time has improved up to 37x and query times by 2-3x versus competitors. New applications, such as real-time public transport flow simulations, will allow for much greater efficiency by enabling agencies/companies to react in advance to congestion or dangerous situations. The performance achieved by this new k-NN acceleration structure will enable new frontiers in crowd simulation.