EdgeVision: Towards Collaborative Video Analytics on Distributed Edges for Performance Maximization

TL;DR

EdgeVision tackles latency and bandwidth challenges in distributed edge video analytics by employing a centralized-training, decentralized-execution MARL framework where autonomous edge nodes collaborate on video preprocessing, DNN selection, and request dispatching. An attention-enhanced actor-critic architecture allows edge agents to learn coordinated policies while distilling useful information from other nodes during training, enabling efficient local decision-making after deployment. Experiments on a four-edge testbed with real-world datasets show EdgeVision achieving significant performance gains (33.6%–86.4% improvement) and substantial frame-drop reductions (up to 92.8%) over baselines, validating the approach's practicality. The work advances edge intelligence by combining collaborative learning, attention-based value estimation, and distributed control to maximize global system performance in dynamic workloads.

Abstract

Deep Neural Network (DNN)-based video analytics significantly improves recognition accuracy in computer vision applications. Deploying DNN models at edge nodes, closer to end users, reduces inference delay and minimizes bandwidth costs. However, these resource-constrained edge nodes may experience substantial delays under heavy workloads, leading to imbalanced workload distribution. While previous efforts focused on optimizing hierarchical device-edge-cloud architectures or centralized clusters for video analytics, we propose addressing these challenges through collaborative distributed and autonomous edge nodes. Despite the intricate control involved, we introduce EdgeVision, a Multiagent Reinforcement Learning (MARL)- based framework for collaborative video analytics on distributed edges. EdgeVision enables edge nodes to autonomously learn policies for video preprocessing, model selection, and request dispatching. Our approach utilizes an actor-critic-based MARL algorithm enhanced with an attention mechanism to learn optimal policies. To validate EdgeVision, we construct a multi-edge testbed and conduct experiments with real-world datasets. Results demonstrate a performance enhancement of 33.6% to 86.4% compared to baseline methods.

Figures (8)

• Figure 1: The system architecture of multi-edge collaborative video analytics. Multiple DNN models are deployed on each edge node, and the edge nodes jointly determine video preprocessing configurations, DNN models for inference, and inference locations to maximize the overall performances.
• Figure 2: The network structure of the actor and critic networks of an agent. The actor network makes control decisions solely based on the local state, while the critic network makes predictions based on global states.
• Figure 3: The convergence of our method under different penalty weights. Our algorithm can converge under different weights.
• Figure 4: The distributions of selected DNN modes for inference and selected resolutions for preprocessing under different weights.
• Figure 5: The average accuracy, delay, request dispatching percentage, and video frame drop percentage under different weights.