分布式智能摄像机协同的非正常事件检测技术研究

A smart camera network has many advantages over traditional visual surveillance systems. Comparing to single camera systems, it can solve some quite difficult questions well, e.g., objects occlusion, objects being observed in too few frames, etc. Comparing with the centralized systems, it will not solely depend on the computational ability of the central computer, and can reduce the bandwidth requirement for transferring the video sequences from a camera to another. Regarding pedestrians and vehicles as the main targets in surveillance systems, the proposed project will focus the research on cooperatively intelligent analysis using multiple smart cameras, and explore the computation models, algorithm mechanisms, and system applications of complicated events detection based on smart camera networks. Key technologies involve the cost-effective feature extraction, the coding method which can efficiently exploit the correlation between the extracted feature data, online unsupervised learning method for constructing and dynamically updating the spatio-temporal correlation models for camera networks, and the methods which infer behaviors by fusing data from multiple smart cameras. In summary, the research of this project is the infrastructure of the smart camera network, and the expected results will also embody the advantages of the system. The research results will help the video surveillance manufacturers acquiring leadership in world competition.

与单摄像机系统相比，智能摄像机网络能很好地解决目标遮挡、目标在监控视频中连续出现的帧数过少等困难；与中心处理监控系统相比，不再过多地依赖中心计算机的处理能力，且能解决传输图像序列需消耗大量带宽的问题。项目将以行人和车辆为主要监控对象，以研究分布式摄像机的智能分析和协同处理为核心，探索以智能摄像机网络实现复杂事件检测的计算模型、算法机理及系统应用。力争在特征提取、与所提取特征具有的相关性相适应的数据压缩编码、摄像机网络时空关联模型的构建和动态更新、综合远景和近景摄像机的复杂事件检测等几个关键技术的研究中取得突破。项目研究构成智能摄像机网络的基础和核心，通过所取得的成果，能真正体现智能摄像机网络作为下一代视频监控系统超越单摄像机系统和中心处理系统的优势，研究成果应用于实践将有助于国内视频监控企业在智能视频监控方面获得领先优势。

项目以行人和车辆为主要监控对象，以多摄像机协同智能分析为应用背景，对一些共性技术进行了研究，具体包括以下几个方面：（1）针对车辆辅助驾驶中的车道偏离和前车防撞等应用，研究了基于深度卷积神经网络的小目标快速检测方法，基于Regionlet特征的车道线检测方法，基于多尺度循环注意力网络的交通标识检测方法；（2）针对以视频为输入的人体姿态和动作分析的应用，研究了基于密集连接注意力金字塔残差网络的人体姿态估计方法，基于动作、多线索结合的视频分类方法；（3）针对摄像机网络中数据压缩传输的应用，研究了一种提高HEVC编码效率的方法；针对低码率传输场合容易产生压缩伪影的问题，研究了一种基于深度神经网络的压缩伪影去除方法。.通过上述研究，本项目在视频目标（特别是小目标）检测、人体姿态分析、视频数据压缩编码和压缩图像恢复增强等领域取得了一些重要成果，发表SCI检索论文10篇；申请专利11项，其中6项已获得授权。