欠覆盖环境下城市多源监控视频大数据高效编码方法研究

Data size of the multi-source urban surveillance video doubles in two years, whilst the video coding efficiency doubles in ten years. The big gap between the increment of the data size and the increment of the coding efficiency leads to huge pressure on data storage. Current joint coding strategy of multi-source videos mainly focuses on the scene redundancy induced by the overlapped area of multiple cameras. However, the global redundancy is caused by the movements of foreground objects between non-overlapped urban surveillance cameras. Therefore, current joint coding strategy will not have high coding efficiency if being directly used on the data from multi-source urban surveillance videos. In order to improve the coding efficiency of non-overlap multi-source urban surveillance video data, we are going to introduce prior knowledge of objects so as to measure the similarity between objects from different videos. Based on that, the composition and distribution of global object redundancy in multi-source surveillance videos are studied. Then, a knowledge-based representation model is proposed, including the common knowledge of objects, postures and factors of the change of image representation. The high level information in the representation models also brings an opportunity for accurate prediction in the joint coding of objects from multiple cameras. Lastly, we are going to develop a unified distortion measurement for parameters of object representation model. Based on that, a rate-distortion model is built to guide the rate allocation of objects among multi-source coding. This project is expected to improve coding efficiency of HEVC by 30% for multi-source coding of surveillance video data, as well as develop. new strategy for the joint coding of multi-source videos.

城市监控视频数据规模两年翻一番，媒体压缩效率十年提高一倍，数据规模增长迅速和压缩效率提升缓慢间的矛盾日益突出。现有多源视频编码方法主要去除摄像机交叉覆盖下的场景冗余，但实际城市监控设备部署主要遵循欠覆盖原则，不满足交叉覆盖假设。城市监控场景下同一运动对象被不同区域监控设备反复摄录会产生大量全局对象冗余。针对这一问题，本项目首先研究全局对象冗余产生机理，引入对象高层语义知识，建立基于知识的跨时空视频间相似性度量方法，揭示多源监控视频中冗余的构成及分布规律。其次，研究对象基于知识的表达模型，在对象预测方法中引入姿态和表观信息，将基于底层特征的局部时空预测扩展到基于高层知识的跨时空预测。最后，研究对象表达模型中不同参数失真的统一度量方法，建立基于对象表观质量的率失真模型，拓展传统码率分配方法，提升编码质量。本项目预期较HEVC单信源独立编码的压缩效率提高30%，探索多源视频联合编码的新方法。

城市监控视频数据规模两年翻一番，媒体压缩效率十年提高一倍，数据规模增长迅速和压缩效率提升缓慢间的矛盾日益突出。现有多源视频编码方法主要去除摄像机交叉覆盖下的场景冗余，但实际城市监控设备部署主要遵循欠覆盖原则，不满足交叉覆盖假设。城市监控场景下同一运动对象被不同区域监控设备反复摄录会产生大量全局对象冗余。针对这一问题，本项目首先研究全局对象冗余产生机理，引入对象高层语义知识，建立基于知识的跨时空视频间相似性度量方法，揭示多源监控视频中冗余的构成及分布规律。其次，研究对象基于知识的表达模型，在对象预测方法中引入姿态和表观信息，将基于底层特征的局部时空预测扩展到基于高层知识的跨时空预测。最后，研究对象表达模型中不同参数失真的统一度量方法，建立基于对象表观质量的率失真模型，拓展传统码率分配方法，提升编码质量。本项目预期较HEVC单信源独立编码的压缩效率提高30%，探索多源视频联合编码的新方法。