| dc.description.abstract |
Green IoV has emerged as a latent solution in the field of autonomous driving for the future Intelligent transportation system (ITS) accompanied with green wireless communication and computational intelligence. It facilitates enhanced traffic management applications, reduced traffic congestion and compatible V2X connectivity. However, GIoV faces significant challenges in providing seamless bandwidth for real-time video analytics, especially under adverse environments, with improved accuracy in autonomous driving. Although deep neural networks (DNN) are effective in locating vehicles, they struggle to frequently access the edge network and maintain accuracy. In addition, their substantial computational demands waste energy and render them infeasible to deploy on resource-constrained devices for low-latency real-time inference. In this paper, we propose a high-speed GIoV (HS-GIoV) framework that models the problem of energy-efficient video analytics accuracy over multiple time-periods using Lyapunov optimization. To solve this problem, we have proposed a novel on-the-fly Traffic Stream Object Detection (TSOD) algorithm which is lightweight and triggers the re-training only when there is an accuracy decline, thereby avoiding unnecessary computations. We have also proposed a heuristic algorithm that solves seamless bandwidth issue using Lagrangian relaxation. We have tested the HS-GIoV on the self-driving kit that comprises NVIDIA high-end devices. It enhances the accuracy around 20% and reduces the training time to approx. 55%. |
en_US |