This paper presents a learning-based model of pedestrian flows that integrates multi scale behaviors such as global route selection and local collision avoidance in urban spaces, particularly focusing on pedestrian movements at Shibuya scramble crossing. Since too much congestion of pedestrian flows can cause serious accidents, mathematically modeling and predicting pedestrian behaviors is important for preventing such accidents and providing a safe and comfortable environment. Although numerous studies have investigated learning-based modeling methods, most of them focus only on the local behavior of pedestrians, such as collision avoidance with neighbors and environmental objects. In an actual environment, pedestrian behavior involves more complicated decision making including global route selection. Moreover, a state transition from stopping to walking at a traffic light should be considered simultaneously. In this study, the proposed model integrates local behaviors with global route selection, using an Attention mechanism to ensure consistent global and local behavior predictions. We recorded video data of pedestrians at Shibuya scramble crossing and trained the proposed model using pedestrian walking trajectory data obtained from the video. Simulations of pedestrian behaviors based on the trained model qualitatively and quantitatively validated that the proposed model can appropriately predict pedestrian behaviors.

翻译：本文提出一种基于学习的行人流模型，该模型整合了城市空间中全局路径选择与局部避碰等多尺度行为，特别聚焦于涩谷交叉路口的行人运动。由于行人流过度拥挤可能引发严重事故，对行人行为进行数学建模与预测对于预防此类事故、提供安全舒适的环境具有重要意义。尽管已有大量研究探讨基于学习的建模方法，但多数仅关注行人局部行为（如与邻近行人及环境物体的避碰）。在实际环境中，行人行为涉及更复杂的决策过程，包括全局路径选择。此外，需同时考虑交通信号灯处从停止到行走的状态转换。本研究提出的模型通过注意力机制整合局部行为与全局路径选择，确保全局与局部行为预测的一致性。我们录制了涩谷交叉路口的行人视频数据，并利用视频提取的行人行走轨迹数据训练所提模型。基于训练模型的仿真实验从定性与定量角度验证了该模型能够准确预测行人行为。