面向实时人群动力学分析的深度基本图网络

当前行人疏散实验中基本图计算方法通常是通过对每个行人进行跟踪实现的。但这种跟踪方法难以实现实时人群动力学分析。针对这一问题,提出了深度基本图网络。实验提出的网络框架由两个模块组成,即多尺度递归卷积神经网络(MSR-Net)和光流模块,分别对行人密度和行人速度进行估计。具体来讲,MSR-Net学习了输入图像与行人密度图之间的映射关系。同时,光流模块实时计算出行人运动速度。基于密度图与速度图的空间对应关系,可以得到基本图。实验表明,我们的方法与传统方法有较好的一致性,而我们的方法可以直接得到行人运动信息,不需要先提取轨迹。同时,基于深度基本图网络还可以实现复杂场景下的异常检测,在人群动力学分析领域具有很好的应用前景。

Deep fundamental diagram network for real-time pedestrian dynamics analysis

Some recent work calculated the fundamental diagram of pedestrian flow by tracking each pedestrian in the crowd from video recordings. However, this method is difficult to realize real-time pedestrian dynamics analysis. To address this problem, this work proposes a novel convolutional neural network based framework, called deep fundamental diagram network, for real-time pedestrian dynamics analysis. Our proposed framework consists of two sub-networks, the multi-scale recursive convolutional neural network (MSR-Net) and optical flow module, accounting for density distribution estimation and pedestrian motion prediction. Specifically, MSR-Net is presented to learn the direct mapping from the input image of pedestrian flow to the output map of crowd density. OF-Net is introduced to predict the velocity and direction of the pedestrian in real-time. In this way, by aligning the position of the pedestrian density map we are able to obtain the fundamental diagram, which shows good agreement with the ones from classical methods but higher computational efficiency. Simultaneously, deep fundamental diagram network can carry out pedestrian anomaly detection, which is meaningful for crowd analysis.