基于局部中层特征的密集场景视频分析研究

With the increase of the crowded activities, the demand for the intelligent analysis of crowded video is urgent for public management and security insurance. Most current video analysis approaches target at low-density people video, but they are not applicaple to chaotic and high-density people scenes. Aiming at the crowded scene video, this research dives into key problems including feature representation, semantic modeling and learning with weakly-supervised data, and proposes a middle level feature based solution. The detailed contents include: 1) with the deep convolutional neural network, proposing a learning-based method to obtain the feature representation for crowded people, utilizing the large available data information via pretraining, to lay a fundation for scene analysis tasks; 2) researching on the people activity local areas mining based on the weakly-supervised learning, the clustering and mining are based on a relational graph, to obtain the middle level units for scena analysis; 3) based on the local activity areas, through the deep learning based population density estimation and multiple features based local activity classification, extracting the middle level semantic features of crowded scenes; 4) based on the middle level features, with contextual modeling and graph-based semi-supervised learning, achieving scene status recognition. Through the above methods, this project is expected to realize a high performance crowded scene video intelligent analysis platform, providing a critical complementarity to conventional low-density people scene video analysis.

密集人群活动越来越多，对其监控视频的智能分析是公共管理、安全保障的迫切需求。现有的视频分析方法普遍针对于中低密度人群视频，对混乱、拥挤的密集人群场景难以适用。本课题针对密集场景视频，研究特征表示、语义模型和利用有限标注数据学习等关键问题，提出基于局部中层特征的研究方案。具体研究内容：1）基于学习的方法提取密集人群特征表示，采用深层卷积神经网络，通过预训练对大量数据信息进行有效利用，为场景分析任务提供基础；2）研究基于弱监督学习的人群活动局部区域自动挖掘，采用关联图聚类挖掘，获取场景分析的语义性中层单元；3）基于局部区域，通过基于深度学习的人群密度估计和融合多特征的局部活动识别，获得场景的中层语义信息；4）基于中层语义特征，通过上下文相关建模和基于关联图的半监督学习，实现密集场景状态识别。通过以上研究可望实现高性能的密集人群场所视频智能分析平台，为传统低密度场景视频分析提供必要补充。

密集人群活动越来越多，对其监控视频的智能分析是公共管理、安全保障的迫切需求。现有的视频分析方法普遍针对于中低密度人群视频，对混乱、拥挤的密集人群场景难以适用。本课题针对密集场景视频，研究特征表示、语义模型和利用有限标注数据学习等关键问题，提出基于局部中层特征的研究方案。具体研究内容：1）基于学习的方法提取密集人群特征表示，采用深层卷积神经网络，通过预训练对大量数据信息进行有效利用，为场景分析任务提供基础；2）研究基于弱监督学习的人群活动局部区域自动挖掘，采用关联图聚类挖掘，获取场景分析的语义性中层单元；3）基于局部区域，通过基于深度学习的人群密度估计和融合多特征的局部活动识别，获得场景的中层语义信息；4）基于中层语义特征，通过上下文相关建模和基于关联图的半监督学习，实现密集场景状态识别。通过以上研究可望实现高性能的密集人群场所视频智能分析平台，为传统低密度场景视频分析提供必要补充。