Non-overlapped Multi-source Surveillance Video Coding Using Two-Layer Knowledge Dictionary

In multi-source surveillance videos, a large number of moving objects are captured by different surveillance cameras. Although the regions that each camera covers are seldom overlapped, similarities of these objects among different videos still result in tremendous global object redundancy. Coding each source in an independent way for multi-source surveillance videos is inefficient due to the ignoring of correlation among different videos. Therefore, a novel coding framework for multi-source surveillance videos using two-layer knowledge dictionary is proposed. By analyzing the characteristics of multi-source surveillance videos in large scale of spatio and time space, a two-layer dictionary is built to explore the global object redundancy. Then, a dictionary-based coding method is developed for moving objects. For any object in multi-source surveillance videos, only some pose parameters and sparse coefficients are required for object representation and reconstruction. The experiment with two simulated surveillance videos has demonstrated that the proposed coding scheme can achieve better coding performance than the main profile of HEVC and can preserve better visual quality.