基于多源视频协同分析的大尺度群体事态理解和预测模型研究

Crowd management has brought enormous challenges to public management, security and safety, as the crowded region, density and size that result from mass activities have been increasing along with the worldwide urbanization and rapid development of modern society and economy. The project studies the understanding and predicting of large-scale crowd events and situations in the distributed cameras networks by integrating the successful achievements in crowd dynamic, computer vision, machine learning and fluid dynamic, which bottoms on the extensive works of the applicants and thorough analysis of occurrence, development and evolution mechanism of crowd behaviors. Specifically, (1) we propose novel algorithms to recognize crowd behaviors based on fluid dynamics and discover rare and abnormal crowd events using active learning strategies, such that critical risks are detected timely, and unexpected accidents are controled before they become serious. (2) We propose to analyze the event topology flow which learns a global visual context to associate partial observations across different camera views, based on which the crowd situations are predicted, such that the derivative and secondary damages are prevented. In general, the main objective of this project is to advance the theories and methods of crowd management through studying the crowd behavior modeling, abnormal discovering, and situation predicting, which is of important theoretical and practical significance in terms of intelligentization and informationization of urban management for our country.

随着我国社会和经济的高速发展，由此带来流动人口增加速度、局部区域人员密集程度及聚集规模都呈现大幅度增加趋势。如果缺乏有效的管控手段，极易发生意外突发事件，导致严重后果。本课题结合申请者过去的工作基础，在深入研究群体事件发生、发展和演化规律的基础上，通过汇聚群体动力学、计算机视觉、机器学习和流体动力学等诸多领域的研究成果，研究基于泛在监控网络的多源协同大尺度群体事态理解和预测模型。本课题(1)提出了基于流体动力学的群体行为识别模型和主动学习框架下的乏样本异常事件捕捉算法，以期实现对临界高危区域检测和突发事件源头控制；(2)通过融合网络拓扑先验和局部视场事件关联得到事件拓扑流，实现群体事态预测，以期望遏制突发事件蔓延，避免衍生和次生灾害。总体来说，本课题将从群体行为建模、异常发现和趋势预测等角度进行研究，力求实现相关理论的突破和创新，对提高我国城市公共安全管理智能化、信息化水平具有重要意义。

本项目着眼于群体目标的行为分析问题，针对群体行为分析中存在的目标众多、模式复杂和彼此干扰等问题，从群体运动模式的表征、群体行为的分析和识别、群体异常行为的发现与识别、以及群体事态趋势的预测等方面开展研究，在充分的借鉴了计算机视觉、机器学习、以及群体动力学等领域最新进展的基础上，从基础理论、关键技术等方面开展了一系列创新性的研究，提出了基于流体动力学分析群体运动模式表征算法、基于社会动力学驱动深度学习算法的群体行为分析、摄像头网络中的目标重识别以及数据与知识融合的深度学习等多种算法，实现了群体目标行为的分析和异常检测。相关工作发表CVPR、AAAI和IJCV等顶级期刊和会议论文24篇，其中CCF A类会议和期刊论文17篇，获得ICME2018铂金最佳论文奖一项，申请国家发明专利8项。本项目培养博士生7人，其中有两人获得了CCF 2019年优秀博士学位论文奖，1人次获得微软学者奖学金，1人次获得百度奖学金。