Effective Task Scheduling for Large-Scale Video Processing

The rapid growth of video surveillance systems has brought the trend of analyzing video objects characteristics for subsequent semantic applications. However, the complexity of extracting object features from surveillance video is substantial due to resource consumption in video transmission and computation in a large-scale distributed environment. Video processing jobs should be adequately assigned to distributed processing servers, without violating the capacity requirement in processing video flows. To resolve this issue, we discuss fundamental design principles for the task scheduling in large-scale video processing systems. We present the architecture and methods of distributing jobs in a resource pool, with considerations on important factors such as the prediction of video flow traffic, the processing workload and the heuristic assignment decision. Proposed methods can be selectively implemented in practical systems with emphasis on satisfying different system requirements.