Langevin Monte Carlo-based Offloading for Edge-Assisted Real-Time Video Analytics

Leveraging edge computing and efficient video inference algorithms facilitates the implementation of various real-time video analytics applications on edge devices. However, existing methods for selecting regions of interest to offload typically only consider the entire video frame, detection bounding boxes on the device, or their aggregated blocks. They offload redundant information or lose critical features of offloaded targets. Furthermore, most current offloading mechanisms either assume the availability of ground-truth prior results or rely on empirical heuristic algorithms, which limits their applicability in real-world scenarios. To address these challenges, this paper proposes an edge-assisted video analysis framework that effectively coordinates detection models of diverse scales between the device and edge server. Additionally, considering the complex dynamics of the system environment, we formulate an offloading problem based on encoding quality and mixed confidence, which aims to maximize the weighted utility of accuracy and end-to-end latency. We propose a Langevin Monte Carlo-based Offloading (LMCO) algorithm, which directly samples from the posterior distribution of historical decision data to accurately model the reward function and select the optimal system parameter configuration. Finally, we evaluate the algorithm’s performance on the real-world video dataset. The experimental results demonstrate that the LMCO outperforms existing baselines across various application scenarios and can quickly provide near-optimal solutions with sublinear regret.