面向复杂公众场景安全监控的异常行为快速检测与识别

The public scene is the important place which involves people daily lives and social activities. The security surveillance problem in the public scene is related to people’s livelihood, and also the social stability. This project begins with the perspective of the video surveillance understanding of complex public scenes, and then focuses on the theory principle and algorithm method about the abnormal activity detection and recognition problem. The severity assessment of the abnormal activity is taken in to account in this project, as well. The major researches in this project are organized as follows. (1) Feature descriptor characterizing the spatio-temporal movement information is studied. Feature compression and selection methods are fussed to extract the feature firstly, and then deep learning is used to mine the extracted feature to provide a high performance feature descriptor. (2) The complex scene which is recorded by the wide-field camera is split into several well-designed sophisticated sub-scene. The activity feature is classified by pattern classification methods. Furthermore, by using the kernel method and sparse representation theory, the abnormal activity is detected efficiently and accurately. (3) Based on the abnormal activity detection result, the severity assessment and activity recognition method are researched. Accordingly, the abnormal activity is evaluated by the assist of fuzzy logics and label distribution learning method. Moreover, the optimized decision is given in the basis of the abnormal activity recognition result. The research results of this project will not only develop the theory of abnormal activity detection and recognition problem in complex public scenes, and will also provide more technical tools in this research area. What is more, the research can provide principal technical support approaches for the realization of smart security surveillance and efficient emergency decision.

公众场景是涉及人们日常生活与社会活动的重要场所，其安全监控事关国计民生与社会稳定。本项目从复杂公众场景之视频监控图像的分析理解入手，研究异常行为的快速检测与识别，及其严重程度的分级预警等关键理论方法和实现算法。主要研究内容为：（1）研究多源视频图像的时空动态特征提取、压缩和优选方法，进而通过深度学习挖掘，以得到高效的深层次特征。（2）针对复杂环境下的大视野监控场景，通过场景区域划分确定典型区域，针对区域行为特征进行分类，并利用核方法优化和稀疏表达数据约减，提出异常行为的快速检测算法。（3）根据异常行为的检测结果，研究异常区域的行为严重性评估与识别方法，进而结合识别结果再辅以模糊逻辑与标记分布学习实现异常行为的快速分级预警，为进一步优化决策提供评判依据。本项目的研究成果将发展和丰富复杂公众场景异常行为快速检测与识别的理论方法和技术工具，可为实现智能化安全监控和高效应急决策提供重要技术支持。

在国家自然科学基金委资助下，课题组围绕基于复杂公众场景视频图像分析理解的特征提取与优选、基于运动行为分类的异常行为快速检测方法、复杂公众场景中的异常行为识别与分级预警这三个主要的研究内容开展研究。针对远距离多摄像头监控的复杂场景，提出了基于时空配准的特征融合方法。在特征表描述方面，本项目提出了多感兴趣区域的协方差的特征描述子，基于卷积自编码的特征描述方法。在异常行为检测方面，提出了基于概率模型深度学习的特征提取与全局与局部异常事件检测算法，变分自编码网络的基础结构为全连接网络，这使得其无法提取输入图片的局部特征，无法充分利用图像的局部相关性的特点。将卷积神经网络与变分自编码结合，充分利用卷积神经网络提取图像局部特征和变分自编码学习输入潜在分布的优势，提升异常行为检测的效果。在分级预警方面，基于自编码的重建误差进行分级，并建立了相关检测数据集与软件模型。针对监控视频情报大数据中的行为识别问题，提出基于深度学习与迁移学习的时空特征挖掘方法。面向小数据集上深度网络模型难以训练的问题，提出了内部迁移学习算法。针对复杂场景中目标运动状态分析的需求，提出了基于局部特征学习的快速实时判别型跟踪方法。搭载摄像头采集设备的无人设备广泛的应用于图像获取与监控方面，针对大范围场景的监控需要，结合现今的无人机的发展，设计了基于多无人机动态相机网络的异常行为检测与目标搜索策略。本项目尝试将理论应用于传感器数据的异常检测、工业产品的异常检测，这些应用的拓展也取得了良好的效果。在本项目的支持下，作为硕士导师带的第一名硕士生乔美娜获得北京航空航天大学十佳研究生、国家奖学金、优秀毕业论文、优秀毕业生，本人获得优秀毕业论文指导教师。发表SCI期刊论文14篇，EI期刊论文3篇，EI会议论文13篇，中文核心期刊论文1篇，教改论文2篇。申请发明专利7项，获得软件著作权1项。