特征自学习机制下的密集群体内多人交互行为异常感知

High crowded scenes not only become the more covert places for illegal activities, but also have the high probability that people in a wider area are quickly affected by abnormal behaviors in a small region. As a result, it is important to find and warn abnormal activities as soon as possible, so that people’s public interests and social security could be well promised. However, monitoring images of high crowded scenes often have problems of heavy occlusions, low resolution of single target and insignificant movements of abnormal activities. So that both algorithms of multiple target tracking and motion flow have their limits to specific crowd’s density or abnormality categories. We would like to build a multi-human abnormal interaction sensing algorithm in dense crowds based on a feature self-learning mechanism. In a fixed viewpoint, probable abnormal areas could be detected based on motion flow analysis. Then the spatio-temporal descriptions about the interaction 3D spaces could be accomplished by multi-layer self-learning sparse features. Associating them with information of human trajectories, spatial distribution and targets’ focusing areas, multi-human interaction recognition and abnormality sensing could be accomplished. The research results not only could improve the development of social abnormal behavior intelligent detection and forewarning, but also could provide information for the research of human-machine interaction, environmental controlling and surveillance, sports and entertainment analysis.

密集群体不仅为违法犯罪事件提供了更加隐秘的实施空间，而且存在小范围异常迅速波及大范围区域的可能。因此，及时发现并预警密集群体中异常事件与民众的公共利益和社会长治久安息息相关。然而，群体视频监控图像普遍存在目标遮挡严重、单目标分辨率低、异常行为运动特征不显著等特点，导致遵循多目标检测跟踪或群体运动流分析思路均局限于特定的群体密度和异常类型。本项目拟建立一种基于特征自学习机制的密集群体多人交互行为异常感知算法架构，在固定视角下，利用密集群体流动性分析获取异常预判区域，再采用多层自学习稀疏特征描述多人互动行为时空区域，并结合行为人的运动轨迹、空间分布、关注区域等信息实现密集群体中的多人交互行为识别和异常感知。本项目的研究成果不仅对公共场所异常行为智能检测及预警，刑侦案件辅助侦破等公共安全事务的智能化发展有积极的推动作用，还能为人机交互,环境控制和监测,体育及娱乐分析等多个领域的应用提供研究基础。

密集群体不仅为违法犯罪事件提供了更加隐秘的实施空间，而且存在小范围异常迅速波及大范围区域的可能。因此，及时发现并预警密集群体中异常事件与民众的公共利益和社会长治久安息息相关。然而，群体视频监控图像普遍存在目标遮挡严重、单目标分辨率低、异常行为运动特征不显著等特点，导致遵循多目标检测跟踪或群体运动流分析思路均局限于特定的群体密度和异常类型。本项目建立了一种基于特征自学习机制的密集群体多人交互行为异常感知算法架构，在固定视角下，首先提出了一种针对中等密度密集场景的小群体实时检测算法。引入基于社会力的目标预测模型对个体运动建模，实现目标方向预测，再结合小群体共有运动状态分析，通过基于目标相关性的滤波算法，完成对密集场景的小群体分类检测。此后，区别于以二维静态图像为对象的传统人脸表情识别，提出一种针对RGB-D图像序列分析的人脸自然表情自动识别算法，以适用于密集群体场景中的低分辨率人脸表情。算法针对预处理后的RGB-D表情图像序列，提取四维时空纹理特征作为局部动态特征；再利用慢特征分析自动检测表情序列的峰值图像，并提取脸部三维几何模型为全局静态特征。结合动、静态特征，经主成分分析降维后输入条件随机场模型完成特征训练和表情识别。该算法不但相较传统静态表情识别算法和其他动态算法具有优越性，而且能够针对自然展现的表情实现自动识别，为今后算法的实用化提供了可能。最后，本项目还提出了一种多模态人体情感感知算法，通过提取视频数据中的时空特征，建立情感语义描述，再采用典型相关分析算法和稀疏保持典型相关分析算法融合面部情绪特征和肢体动作情绪特征实现最终情感感知，有效提高了密集场景下的人体情感认知算法的有效性。本项目的研究成果不仅对公共场所异常行为智能检测及预警，刑侦案件辅助侦破等公共安全事务的智能化发展有积极的推动作用，还能为人机交互,环境控制和监测,体育及娱乐分析等多个领域的应用提供研究基础。