ILS-YOLO: An Improved List Scheduling Algorithm for YoloLite DAGs in Vehicular Edge Computing

Abstract

Real-time vehicle detection is critical for intelligent transportation systems, yet deploying lightweight models like YoloLite on resource-constrained vehicular units leads to prohibitive latency. To address this challenge, this paper introduces ILS-YOLO, an Improved List Scheduling algorithm for YoloLite Directed Acyclic Graphs (DAGs) in Vehicular Edge Computing (VEC). The algorithm models the YoloLite inference pipeline as a DAG of sub-tasks and introduces two key innovations: a Unified Scheduling Table to mitigate resource fragmentation on multi-core edge servers and a Transfer Scheduling Table to accurately model link contention and communication overhead. Performance is evaluated through extensive simulations using synthetic DAGs with varying structures (up to 200 nodes) and Communication-to-Computation Ratios (CCRs), alongside a case study using parameters derived from a complex perception task and real-world traffic data. Results show that ILS-YOLO significantly reduces end-to-end detection latency, achieving a speedup of up to 2.5× over baseline scheduling heuristics for large, computation-intensive DAGs (n=200, low CCR). This work presents a robust and efficient scheduling solution that makes low-latency, high-accuracy vehicle detection feasible in practical VEC environments.