GPU accelerated Nature Inspired Methods for Modelling Large Scale Bi-Directional Pedestrian Movement

Pedestrian movement, although ubiquitous and well-studied, is still not that well understood due to the complicating nature of the embedded social dynamics. Interest among researchers in simulating pedestrian movement and interactions has grown significantly in part due to increased computational and visualization capabilities afforded by high power computing. Different approaches have been adopted to simulate pedestrian movement under various circumstances and interactions. In the present work, bi-directional crowd movement is simulated where an equal numbers of individuals try to reach the opposite sides of an environment. Two movement methods are considered. First a Least Effort Model (LEM) is investigated where agents try to take an optimal path with as minimal changes from their intended path as possible. Following this, a modified form of Ant Colony Optimization (ACO) is proposed, where individuals are guided by a goal of reaching the other side in a least effort mode as well as a pheromone trail left by predecessors. The basic idea is to increase agent interaction, thereby more closely reflecting a real world scenario. The methodology utilizes Graphics Processing Units (GPUs) for general purpose computing using the CUDA platform. Because of the inherent parallel properties associated with pedestrian movement such as proximate interactions of individuals on a 2D grid, GPUs are well suited. The main feature of the implementation undertaken here is that the parallelism is data driven. The data driven implementation leads to a speedup up to 18x compared to its sequential counterpart running on a single threaded CPU. The numbers of pedestrians considered in the model ranged from 2K to 100K representing numbers typical of mass gathering events. A detailed discussion addresses implementation challenges faced and averted.

💡 Research Summary

The paper presents a GPU‑accelerated framework for simulating large‑scale bi‑directional pedestrian movement, where equal numbers of agents travel toward opposite sides of a confined environment. Two nature‑inspired algorithms are investigated. The first, a Least Effort Model (LEM), directs each agent to follow a path that deviates minimally from a straight line toward its goal, using a cost function based on angular deviation, proximity to other agents, and collision risk. The second approach adapts Ant Colony Optimization (ACO): agents deposit a pheromone trace as they move, and subsequent agents choose their next cell by balancing the pheromone intensity with the same least‑effort cost. Pheromone evaporation is modeled to allow outdated routes to fade, thereby enabling dynamic adaptation of the crowd flow.

Implementation leverages NVIDIA’s CUDA platform with a data‑driven parallelism strategy. Each pedestrian is mapped to a single CUDA thread; threads are organized into blocks that correspond to spatially adjacent regions, allowing the use of shared memory for rapid access to neighboring cell information. To avoid race conditions during pheromone updates, the authors separate reading and writing phases and employ lightweight synchronization primitives. Atomic operations are minimized, and memory accesses are coalesced through padding and warp‑level optimizations. This design yields up to an 18× speed‑up over a sequential CPU baseline when running on a single‑threaded core.

Experiments cover population sizes from 2 000 to 100 000 agents, reflecting typical mass‑gathering scenarios. Results show that while both models maintain overall flow and avoid deadlocks, the ACO‑enhanced version achieves higher directional alignment and lower average travel times (approximately 12‑25 % improvement) especially at higher densities. Sensitivity analyses on pheromone decay rates and the weighting of the least‑effort term reveal that appropriate tuning is crucial for balancing individual efficiency against collective coordination.

The authors also discuss implementation challenges such as memory contention, thread divergence, and data transfer overhead. Solutions include memory padding, warp‑level pre‑emptive scheduling, and overlapping computation with data movement via asynchronous streams. The paper concludes that the combination of biologically inspired decision rules and GPU‑based data‑driven parallelism provides a scalable, high‑fidelity tool for crowd simulation, with potential applications in urban planning, emergency evacuation modeling, and immersive virtual environments.