Migratory Perception in Edge-Assisted Internet of Vehicles

Abstract

Listen

Translate article

Autonomous driving technology heavily relies on the accurate perception of traffic environments, mainly through roadside cameras and LiDARs. Although several popular and robust 2D and 3D object detection methods exist, including R-CNN, YOLO, SSD, PointPillar, and VoxelNet, the perception range and accuracy of an individual vehicle can be limited by blocking from other vehicles or buildings. A solution is to harness roadside perception infrastructures for vehicle–infrastructure cooperative perception, using edge computing for real-time intermediate features extraction and V2X networks for transmitting these features to vehicles. This emerging migratory perception paradigm requires deploying exclusive cooperative perception services on edge servers and involves the migration of perception services to reduce response time. In such a setup, competition among multiple cooperative perception services exists due to limited edge resources. This study proposes a multi-agent reinforcement learning (MADRL)-based service scheduling method for migratory perception in vehicle–infrastructure cooperative perception, utilizing a discrete time-varying graph to model the relationship between service nodes and edge server nodes. This MADRL-based approach can efficiently address the challenges of service placement and migration in resource-limited environments, minimize latency, and maximize resource utilization for migratory perception services on edge servers.