大型公共场所群体异常行为跨域感知与安全事件预警

There exists a pressing need for the next-generation prevention technology to handle the grand threats from social security incidents in public places in China. This research proposal aims at the fundamental problems with early warning of group security incidents in public places being inspected by video surveillance. The technical bottlenecks include abnormal behavior detection, panoramic perception, and evolutionary analysis of critical events: to explore the joint recognition method for individual and group behaviors to promote the poor performance of group abnormal behavior identification in large public places; to develop the cross domain correlation method to enable panoramic perception of group abnormal behaviors which was previously impossible under an individual fixed surveillance scene; to realize the evolutionary analysis for critical security events by fostering a DDDAS method empowered by a multi-scale crowd behavior model. The main objective is to achieve the capability of comprehensive understanding of the life-cycle of abnormal behaviors and the consecutive security events. We aim to significantly enhance the intelligence for the modeling and simulation system of large-scale public places, and the following specifications are expected: (1) online simulation analysis of tens of thousands people; and (2) more than 80% accuracy for the prediction of abnormal behaviors and events. The anticipated outputs will (1) produce fundamental theories for handling the group security events by enabling spatial and temporal evolution analysis on their life cycles; and (2) propel the quick development of interdisciplinary research area upon video analysis, intelligent computing, and social security.

大型公共场所群体安全事件防范属社会安全领域的国家重大需求。本项目面向视频监控条件下大型场所群体安全事件预警的基础科学问题，重点突破异常发现、全景感知、演化预警三个方面的技术瓶颈：针对大型场所群体异常行为辨识性能低下的问题，研究个体与群体异常行为联合辨识方法；针对单一固定监控场景无法全程感知群体异常行为的难题，研究基于跨域关联的群体异常行为全景感知方法；针对异常行为与安全事件全生命周期演化过程理解能力缺乏的难题，研究数据与模型混合驱动的复杂群体安全事件变尺度演化预测方法。通过上述工作，实质性提升群体异常行为识别的准确率, 获得群体异常行为的全景感知能力，实现群体安全预警分析能力的智能增强，具备万人以上规模场景在线仿真分析能力，异常行为与事件演化预测的正确性提高到80%以上。预期研究成果将为群体安全事件的全生命周期时空演化分析提供理论支持，推动视频分析、智能计算和社会安全等领域的交叉学科发展。

面向监控条件下大型场所群体安全事件预警的基础科学问题，重点突破异常发现、全景感知、演化预警三个方面的技术瓶颈：针对大型场所群体异常行为辨识性能低下的问题，研究个体与群体异常行为联合辨识方法；针对单一固定监控场景无法全程感知群体异常行为的难题，研究基于跨域关联的群体异常行为全景感知方法；针对异常行为与安全事件全生命周期演化过程理解能力缺乏的难题，研究数据与模型混合驱动的复杂群体安全事件变尺度演化预测方法，实质性提升个体到群体异常行为识别的准确率，实现群体安全预警分析能力的智能增强。具体而言，研究监控场景路径发现算法，发展移动监控自主路径跟随技术框架，实现未知环境中多智能体协同全景导航；通过监控时间序列数据分析发展区域异常行为模式发现算法，发展监控对象注意力估计与情绪分析方法，建立视频监控条件下异常个体跟踪机制；提出时间序列大数据特征构建与异常预测机制，研发复杂群体场景演化预测技术。其中时间序列大数据处理分析系列方法，已成功推广到生理信号分析等领域。在IEEE TKDE/TSC/TMI, ACM CUSR等信息科学及交叉学科权威期刊发表SCI论文19篇，包括中科院认定Top期刊8篇、中国计算机学会认定A类（CCF-A）期刊论文1篇/CCF-B类5篇、中科院SCI一区论文3篇/二区论文6篇、ESI高被引论文1篇；发表EI会议论文2篇，包括CCF-A会议论文1篇；发表CCF-A中文期刊论文1篇；新增学术论文共计22篇。授权发明专利3项、审查中1项。荣获湖北省科技进步一等奖1项（序3）、全国研究生人工智能大赛一等奖1项。