Abstract Pedestrian traffic is an important subject of surveillance to ensure public safety and traffic management, which may benefit from intelligent and continuous analysis of pedestrian videos. State-of-the-art methods for intelligent pedestrian surveillance have a number of limitations in automating and deriving useful information of high-density pedestrian traffic (HDPT) using closed circuit television (CCTV) images. This work introduces an automatic and improved HDPT surveillance system by integrating and optimizing multiple computational steps to predict pedestrian distribution from input video frames. A fast and efficient particle image velocimetry (PIV) technique is proposed to yield pedestrian velocities. A machine learning regressor model, boosted Ferns, is used to improve pedestrian count and density estimation: an essential metric for HDPT analysis. A camera perspective model is proposed to improve the speed and position estimates of HDPT by projecting 2D image pixels to 3D world-coordinate dat. All these functional improvements in HDPT velocity and displacement estimations are used as inputs to a sophisticated pedestrian flow evolution model, PEDFLOW to predict HDPT distribution at a future time point, which is a crucial information for pedestrian traffic management. The predicted and simulated HDPT properties (density, velocity) obtained using the proposed framework show low errors when compared to the ground truth data. The proposed framework is computationally efficient, suitable for multiple camera feeds with HDPT videos, and capable of rapidly analyzing and predicting flows of thousands of pedestrians. The paper shows one of the first steps towards fully integrated CCTV-based automated HDPT management system.

中文翻译：

---

一种新的高密度行人流自动监测和分析框架

摘要 行人交通是保障公共安全和交通管理的重要监控对象，对行人视频进行智能、连续的分析可能会使其受益。最先进的智能行人监控方法在使用闭路电视 (CCTV) 图像自动化和获取高密度行人交通 (HDPT) 的有用信息方面存在许多局限性。这项工作通过集成和优化多个计算步骤来从输入视频帧预测行人分布，从而引入了一种自动和改进的 HDPT 监控系统。提出了一种快速有效的粒子图像测速 (PIV) 技术来产生行人速度。机器学习回归模型，boosted Ferns，用于改进行人数量和密度估计：HDPT 分析的基本指标。提出了一种相机透视模型，通过将 2D 图像像素投影到 3D 世界坐标数据来提高 HDPT 的速度和位置估计。HDPT 速度和位移估计中的所有这些功能改进被用作复杂的行人流演化模型 PEDFLOW 的输入，用于预测未来时间点的 HDPT 分布，这是行人交通管理的关键信息。与地面实况数据相比，使用所提出的框架获得的预测和模拟的 HDPT 属性（密度、速度）显示出较低的误差。所提出的框架计算效率高，适用于带有 HDPT 视频的多摄像头馈送，并且能够快速分析和预测数千行人的流量。