多源视频准确群体检测方法

Group detection is meaningful to in a log of fields, including public security, public traffic and so on. Crowd video of a single viewpoint, which is short of complete spatial information of group behavior, cannot fulfill the performance requirement of distinguishing members from different group. Multi-sourced video, noted as videos captured by cameras from various view angles or locations are capable of supplying even more spatial-temporal information on group behavior. However, the attribute of coupling and complementary of spatial-temporal group.information derives the problems of attribute profiling for group behavior respect to both incompleteness and inconsistency. This problem is marked as the difficulty of merging group information from multi-soured video to improve the accuracy of group detection. In this proposal, we adopt the semantic algebra and event graph model widely adopted in the field of the Internet of Things, so as to study the merging of multi-level semantic information of multi-sourced video groups with respect to three aspects: 1) Build up semantic algebra model and complex event.processing model for group and group events from multi-sourced videos. 2) Firstly, study the complementary of attribute profiling among groups, including group detection algorithms based on matching and fuzzy inference over groups’ attribute profiling. Secondly, study the coupling of group event, including group detection algorithm based on dynamic merging and path prediction of groups’ event graph. By.introducing complex event processing model on the high-level semantic information of video groups, we propose to match and merge the spatial-temporal information of groups captured in multi-sourced videos. It will be derived as the novel multi-sourced video group detection method with high accuracy and promote the development and application of key technologies in intelligent video analysis.

群体检测在城市公共安全，公共交通等领域具有重要意义。单一视角视频缺乏群体行为完整时空信息，不足以完全区分不同群体个体间的差异。多角度、多位置的视频设备采集的多源视频捕获的群体信息具有耦合和互补性。然而，多源视频间群体行为的属性描述存在不完整和不一致问题，成为融合多源视频群体信息进行群体检测，进一步提高群体检测准确度的难点。本课题拟引入物联网应用中广泛应用的复杂事件处理模型，从模型和算法方面研究多源视频群体行为多层次语义信息融合，1）建立多源视频群体、群体事件的语义代数和复杂事件处理模型；2）首先，挖掘群体属性描述间互补特性，研究群体属性匹配和模糊推理群体检测算法；然后，挖掘群体事件图间耦合特性，研究群体事件事件图动态融合和路径预测群体检测算法。从而，通过研究群体高层语义信息复杂事件模型，匹配和融合多源视频群体空间时序信息，形成多源视频群体准确检测新方法，推动智能视频分析关键。

群体检测在城市公共安全，公共交通等领域具有重要意义。立足于群体行为分析领域，本项目从视频分析和编码过程出发，研究面向监控视频中群体准确发现的智能视频编码和分析方法。包括：1）鉴于群体的瞬时一致运动方向对于由于运动个体跟踪和视频拍摄角度带来的运动特征误差具有鲁棒性，有助于提高群体运动一致性，从而提高密集场景中一致运动群体检测性能，提出一种与群体运动一致性检测方法无关的群体运动一致性过滤方法，提高群体中个体间的运动一致性，帮助群体运动一致性检测方法进行群体运动一致性方向判断，进而提高群体运动一致性。2）面向群体行为识别研究智能视频编码方法，初步形成面向机器视觉的智能视频编码方法以及高效视频编码方法，从而提高视频群体检测的准确性。3）面向多源视频群体行为分析探索了立体图像质量评价，充分研究了双目特征点匹配特性，为多源视频间群体属性匹配奠定了基础。本项目的研究通过提高视频编码效率，从而帮助准确检测出同一群体出现的所有视频位置信息，提高视频群体检测精度，推动智能视频分析关键技术发展，满足公共安全和公共交通等领域应用需求 。