Data-Driven Windows to Accelerate Video Stream Content Extraction in Complex Event Processing

This work presents a data-driven adaptive windowing approach to accelerate video content extraction in DNN-based Complex Event Processing (CEP) systems. The CEP windows continuously monitor low-level content of incoming video frames and exploit interframe correlations to accelerate the overall DNN content extraction process. The two main contributions are: 1) technique to create micro-batches of similar frames within the window by measuring dissimilarities among them, and 2) optimal frame resolution within micro-batches under specified accuracy thresholds for fast model processing. The initial experimental results show that our adaptive micro-batching approach improves 3.75X model throughput execution while maintaining application-level latency bounds under required accuracy constraints.