Abstract: Abstract: To address the problem that traditional evacuation models using third-person global views fail to reflect visual occlusion and perspective scaling, making it difficult to truly reproduce vision-dependent pathfinding behaviors in complex scenarios, a first-person vision-driven pathfinding model for evacuation in complex scenarios was proposed. First, a Dynamic Scanning Algorithm (DSA) was utilized to acquire first-person visual information in real time. Then, in the upper-level navigation decision-making, a Proximal Policy Optimization (PPO) algorithm was adopted as the core to output relay waypoints based on the first-person visual information. Finally, in the lower-level motion model, a collision-free speed model was applied to control agents to achieve local obstacle avoidance while tracking the waypoints. Simulation results show that compared with the traditional end-to-end PPO direct control model using global state input, the time steps consumed by the proposed model to reach convergence are reduced by 90%. In asymmetric exit scenarios, about 50% of the evacuation traffic is borne by the distant exits, and the core interval of the total evacuation time of agents is highly consistent with the 3~10 seconds of the real baseline experiment, which verifies that the proposed model can successfully reproduce complex behaviors such as navigation in unfamiliar environments and congestion avoidance. The theoretical modeling of pathfinding decision-making mechanisms under limited fields of view is deepened, and the proposed model can be used as the underlying engine of auxiliary decision-making systems to quantitatively diagnose visual blind spots in buildings and evaluate the effectiveness of emergency evacuation signs.

Key words: emergency evacuation, crowd simulation, first-person vision, dynamic scanning algorithm, proximal policy optimization algorithm, reinforcement learning, collision-free speed model

摘要： 摘 要: 针对传统疏散模型采用第三人称全局视野导致无法体现视觉遮挡与透视缩放，进而难以真实复现复杂场景下视觉依赖型寻路行为的问题，提出一种第一人称视野驱动的复杂场景疏散寻路模型。首先，利用动态扫描算法（DSA）实时获取第一人称视野信息；其次，在上层导航决策中，以近端策略优化（PPO）算法为核心，基于第一人称视野信息输出中继导航点；最后，在下层运动模型中，改进优化速度模型控制智能体在追踪导航点的同时实现局部避障。仿真实验结果表明：相较于采用全局状态输入的传统端到端PPO直接控制模型，所提模型收敛消耗的时间步减少了90%；在非对称出口场景中远端出口承担了约50%的疏散流量，智能体总疏散时间的核心区间与真实基准实验的3~10秒高度吻合，验证了该模型能够成功复现陌生环境寻路、拥堵规避等复杂行为。本研究深化了有限视野下寻路决策机制的理论建模，并可作为辅助决策系统的底层引擎，用于量化诊断建筑内的视觉盲区及评估应急疏散标识的有效性。

关键词: 应急疏散, 人群仿真, 第一人称视野, 动态扫描算法, 近端策略优化算法, 强化学习, 优化速度模型