Kite: Link-Adaptive and Real-Time Object Detection in Dynamic Edge Networks

Vision-based real-time object detection has become a key fundamental service for smart-city applications such as auto-drive and digital twins. Due to the limited resource available at camera devices, edge-assisted object detection has attracted increasing research attention. The existing edge-assisted schemes often assume stable or averaged wireless links during the frame offloading process. However, the assumption does not hold in real-world dynamic edge networks and will lead to significant performance degradation in terms of both detection latency and accuracy. In this paper, we propose $Kite$, a link-adaptive scheme for real-time object detection. Based on measurement studies and systematic analysis, we devise a lightweight yet representative performance indicator – “frame-anchor” distance, to incorporate the immeasurable impact of wireless dynamics into a measurable metric. Based on this performance indicator, we model the offloading process as an integer nonlinear programming problem, and propose an online link-adaptive algorithm for frame offloading decisions. We implement $Kite$ in a neuro-enhanced live streaming application and conduct comparative experiments with four different datasets in WiFi/LTE based edge networks. The results show that Kite can improve the detection accuracy by 40.53% in highly dynamic networks, compared to the state-of-the-art works.