A Cloud-Edge Framework for Video Stream Analysis using Macroblock Partitioning

Deep neural networks (DNNs) have become the state-of-the-art solution for video stream analysis, but their high delay poses a challenge in meeting real-time requirements. Existing research focuses on compressing video while retaining enough information for server-side DNNs to perform high-precision inference. However, this approach can only reduce network transmission delay and may not be optimal for massive video analysis requests. This paper proposes a Cloud-Edge collaborative Video Analysis (CEVA) framework that uses depth estimation technology to divide the inference difficulty of different macroblocks in the video and offload them to edge and cloud servers for inference. Experimental results on the VisDrone dataset demonstrate that CEVA outperforms server-side feedback-driven methods in inference accuracy and reduces waiting delay by an average of 264 milliseconds. Compared to traditional offloading methods, CEVA doubles accuracy and increases the average delay by less than 100 milliseconds.