Platform-Adaptive High-Throughput Surveillance Video Condensation on Heterogeneous Processor Clusters

摘要

Directly browsing and analyzing numerous surveillance videos is inefficient for human operators. Video condensation is a technical solution to fast video browsing. On the one hand, traditional video condensation methods that skip frames using simple strategies may lose some important frames. On the other hand, the methods that rearrange frame contexts improve the browsing efficiency, but are not easy to be accelerated using the data processing centers with various hardware configurations. In this paper, we propose a platform-adaptive video condensation system based on change detection, which is easy to accelerate and keeps important frames accurately. To take full advantage of hardware acceleration, we implement each module of the proposed system using multithreading and GPU acceleration, and then further accelerate the system by exploiting the task-level parallelism. We solve the computational resources assignment problem via local search method. To be platform-adaptive, the combination of module using different hardware acceleration are compared to choose the optimal combination to make full use of the computational resources. Detailed experiments are conducted to validate the accuracy of the proposed system, the efficiency of the platform-adaptive mechanism and the high throughput performance.