项目名称: 通用时序逻辑表达下的视频时空行为理解研究与应用
项目编号: No.61502006
项目类型: 青年科学基金项目
立项/批准年度: 2016
项目学科: 计算机科学学科
项目作者: 郑爱华
作者单位: 安徽大学
项目金额: 20万元
中文摘要: 传统的视频分析方法在视频的时序信息表达上缺乏明确的表述,这使得视频中丰富的时序关系无法得到充分的挖掘。本项目在通用时序逻辑表达描述视频事件的基础上,主要研究包括特征提取与运动表征、行为识别、高层行为理解几个基本内容的行为理解。针对现有事件表征方法仅仅描述了空间特征和简单的时序特征,本项目研究包含丰富的时序信息的时空特征表征,并针对事件本身具有模糊性的特点,研究基于软聚类的运动轨迹学习方法,将轨迹划分到多个可能的事件类中。在此基础上根据轨迹的时空特征,通过时序模式挖掘来实现无需事先标定的并能处理空事件的时序建模的方法,来挖掘视频行为中的丰富的时序结构和时序依赖性,并以此来增强空间识别效果。最后采用概率模型的方法,完成随意拓扑结构的多相机网络复杂场景下、无需轨迹匹配的行为理解方法,最终实现时空两个层面的异常检测。
中文关键词: 通用时序逻辑表达;行为理解;时空分析
英文摘要: Traditional methods of video analysis lack of explicit representation of temporal information, which makes the rich temporal relationship couldn't be fully explored. Based on the general expression of temporal logic for video event description, this proposal mainly research the task of behavior understanding including feature extraction and motion representation, behavior recognition and high-level behavior understanding. In view of the existing characterization methods considering only spatial and simple temporal features, this proposal researches the spatial-temporal characterization which contains the rich temporal information. Thinking that events are always ambiguity themselves, this proposal studies a soft clustering method to learn the generated trajectories and then divide them into multiple possible event classes. On this basis, temporal pattern is modeled for video event with empty time-series and without calibration though temporal pattern mining, then the rich temporal structure and temporal dependencies among behaviors are mined to enhance the spatial detection. Finally, the probabilistic model is adopted to achieve correspondence-free behavior understanding with complex scene in multi-camera networks with arbitrary topology and eventually achieve the abnormal detection in both spatial and temporal level.
英文关键词: General Temporal Logic Expression; Behavior Understanding;Spatial-Temporal Analysis