Simulating how crowds move through complex environments is essential for applications in urban planning, robotics, gaming, and safety analysis. However, many real-world spaces—such as multi-story buildings, staircases, and layered architectural designs—are too complex for traditional 2D or CPU-based crowd simulation methods, which often oversimplify geometry or become computationally infeasible. This thesis introduces a fully GPU-accelerated crowd simulation framework that efficiently handles complex, multi-layered 3D environments. Building on the Continuum Crowds algorithm, our method extracts walkable surfaces at different height levels, identifies vertical obstructions, and enables real-time navigation for thousands of agents. The system avoids common artifacts such as unrealistic wall clipping and enables realistic movement across layers. In performance benchmarks, our GPU implementation achieves a speedup between 5× and 1000× over the original CPU-based method, depending on scenario complexity, demonstrating strong scalability and consistent real-time performance. This makes it particularly valuable in domains such as robotics, where anticipating pedestrian flow is crucial for safe and intelligent robot navigation in dynamic environments.