Parallelization and optimization of the walkable area mesh generating pipeline

Abstract

Modern crowd simulation is performed by using a navigation mesh to allow agents to plan paths in the environment. By automatically generating a navigation mesh from the 3D environment the walkable area, i.e. the area agents can navigate on, can be extracted. In the paper by A. Hillebrand and a thesis by M. Polak a filtering pipeline was proposed that extracts the walkable areas from an input environment by applying filters sequentially. For non-complex environments, this yielded low execution times and exact navigation meshes. However for complex, large environments this approach struggles to perform well in terms of execution times and memory consumption. Consequently, only small 3D environments with low numbers of polygons can be handled in practice.

This thesis proposes an extension on the filtering pipeline, which exploits the parallel nature of current computers and networks, to tackle these problems. A parallel filtering pipeline is introduced which subdivides the original environment into smaller partitions and distributes these over available resources (threads and/or computing nodes) to solve the walkable area extraction process in a parallel and/or distributed environment. Several partitioning methods are proposed and explored to find optimal load balancing and low execution times on big and/or complex environments.

Additional methods (and their shortcomings) of gaining low execution times are also explored, such as rounding coordinates. The problems with rounding exact coordinates to non-exact coordinates are explored thoroughly. Additionally, algorithms (or solutions) are proposed to grant compatibility between an exact environment and rounded 3D data. We explore the scalability, the speed-ups, the efficiency and the correctness of the proposed parallel pipeline.

For large environments, the speed-ups are substantial of up to 15× faster than the original sequential pipeline. For small environments the speed-ups are less impressive of up to 2× faster. The proposed pipeline scales consistently well with up to 8 cores and provides a mesh comparable in quality to the original pipeline.