Fine-grained spatial-temporal compression for edge-device collaborative video analytics

Driven by the prosperity of edge computing and video analysis algorithms, edge video analytics have been widely deployed to enable real-time decision-making. However, as the number of edge device increases and intelligent algorithms become more complex, the limited bandwidth resources are struggling to meet the stringent response time requirement of many real-time applications. In this paper, we propose a spatial-temporal fine-grained compression scheme for the edge-device collaborative video analytics system, which utilizes the spatial-temporal characteristics of video data to perform differentiated compression on various partitions to reduce the end-to-end inference latency. Moreover, we formulate an optimization problem for joint compression and offloading decision to maximize detection accuracy while minimizing inference latency. Considering the temporal correlation in the optimization problem and the complex coupling between variables, the Dueling Double Deep Q-Networks (D3QN) algorithm is adopted for solution. Experiments results show that compared to the baselines, our proposed scheme achieves higher accuracy while effectively reducing inference latency.