Towards Efficient Autonomous Vehicle Systems: A Multi-Layer Approach

Abstract

This dissertation presents a multi-layer framework to enhance the efficiency, predictability, and scalability of autonomous vehicle (AV) systems by addressing critical challenges across system, application, and communication layers. At the system layer, a novel Extended Conflict Directed Graph (ECDG) model is introduced to enable dynamic lane assignment in unsignalized intersection management, improving flexibility and traffic flow coordination. A scheduling algorithm based on breadth-first search achieves up to 16.3% reduction in intersection evacuation time and 27.7% improvement in traffic efficiency. At the application layer, a 1-opt local search-based scheduling framework is proposed for non-preemptive real-time Directed Acyclic Graph (DAG) tasks. Using convex subspace partitioning and linear programming to minimize timing metrics such as Data Age, Reaction Time, and Time Disparity, the approach yields 20% to 40% reductions in worst-case latency with strong scalability and polynomial-time complexity. At the communication layer, the Partitioned Combined-DBP-TCCP (PCDT) protocol enables wait-free multicore data communication with configurable buffer strategies tailored to task timing requirements. Two optimization strategies, priority assignment with preemption thresholds (PA-MBTT) and wait-free-aware task partitioning (WFAP+), further reduce memory demands, achieving over 50% savings in simulation and 43% in a real-world automotive case study. Together, these contributions form an integrated and theoretically grounded solution for building real-time, safe, and resource-efficient AV platforms, advancing the state of the art in autonomous systems and real-time embedded computing.