Currus - A Compound AI Approach to Distributed Vehicle Trajectory Reconstruction in the Edge-Cloud

Abstract

Increasing traffic volume poses significant challenges for large cities. Advances in different fields of Artificial Intelligence (AI) enable the recognition of vehicles on traffic camera snapshots and the reconstruction of vehicle trajectories, which present a valuable resource for road and traffic planning. However, high traffic volume also yields a high volume of vehicle data, which requires powerful cloud resources for processing, resulting in high costs, considerable processing delays or data volume limits in the absence of these resources. In this paper, we present Currus, a distributed Vehicle Trajectory Reconstruction (VTR) system that leverages Compound AI, distributes the computational load to multiple city regions, which can be handled by cheaper edge servers, and relies on the cloud only for merging the regional results. Currus trades a slight drop in accuracy to process 3X the amount of traffic data of its centralized version, while being up to 24% faster than its centralized version.