A Unified Software-Defined Autonomous Vehicle Network and Urban Congestion Prediction Method

Abstract

Urban traffic congestion is worsening and accurate traffic congestion prediction is essential to address this issue. Current studies mainly concentrate on manned vehicles, overlooking the burgeoning traffic flow that includes both manned and autonomous vehicles. While road infrastructures and autonomous vehicles could alleviate congestion through information exchange, current infrastructure and vehicle diversity hinder effective data collection and management. This paper proposes a unified Software-Defined Autonomous Vehicle Network (SDAVN) to consistently compute traffic parameters such as average velocity, traffic flow, and occupancy using real-time mobility data from autonomous vehicles and connected manned vehicles. Additionally, we propose an effective SDAVN congestion prediction method featuring a Transformer-based traffic parameter prediction module and a congestion detection module employing an extended Spatio-Temporal Self-Organizing Mapping (STSOM). We optimize the 2D SOM to a 3D model to learn more effectively spatio-temporal characteristics. Furthermore, we introduce an asymmetric loss function to address the imbalance between congested and uncongested samples. Experimental results demonstrate the superior long-term congestion prediction performance of our method compared to existing approaches at both road and lane levels across traditional traffic datasets and simulations of real automated driving environments.