A Zero-Copy VPU-assisted Edge–Cloud Computing Framework for bandwidth-efficient video analytics

Large-scale intelligent surveillance applications impose increasing pressure on edge–cloud infrastructures because continuous high-resolution video must share limited uplink bandwidth and edge compute resources. We present an edge–cloud video-analytics prototype that moves spatial cropping and discrete temporal sampling into the Direct Memory Access (DMA) path of a custom Video Processing Unit (VPU), followed by hardware H.264 encoding and cloud-side YOLOv8 inference. We re-evaluate the pipeline on four MOT16 training sequences using fixed-quantization (CQP) encoding and an explicitly defined person-detection AP50 protocol. Temporal sampling reduces payload by 66.09% on average, while macro AP50 changes from 0.6029 to 0.5154; the per-scene AP50 change ranges from −0.0247 to −0.2122. A fixed crop is substantially less robust, with a macro AP50 decrease of 0.1795 and severe failures on MOT16-04 and MOT16-10. On MOT16-04, the joint crop-and-sample configuration reduces payload by 81.05% but lowers AP50 from 0.6263 to 0.1792. An archived software–VPU timing pair shows a 97.49% reduction in process CPU time (72.697 to 1.827 s), although wall time increases from 6.600 to 8.799 s; a current full-frame run consumes 9.0% of one logical CPU. Active two- and eight-card tests expose a shared DMA-mapping failure rather than linear scaling. These measurements establish both the bandwidth benefit and the scene-, motion-, and driver-dependent boundaries of the current prototype.