ISI Journals

Autonomous driving has emerged as one of the most challenging yet transformative domains in artificial intelligence, requiring the integration of robust perception workflows that can process massive streams of sensory data in real time. The complexity of these workflows is heightened by the computational burden of deep learning models, which often demand significant hardware acceleration and software-level optimization to achieve operational safety and efficiency. Traditional computing pipelines in perception systems encounter bottlenecks due to high latency, inefficient GPU utilization, and suboptimal inference scheduling. To address these challenges, a growing body of research has emphasized deep learning optimization, GPU-based acceleration, and the integration of software- defined vehicle (SDV) architectures. These innovations enable the scaling of perception modules for real-time object detection, semantic segmentation, motion prediction, and sensor fusion across diverse driving environments. This article investigates efficient perception workflows in autonomous vehicles, focusing on deep learning optimization strategies that reduce latency, enhance real-time inference, and improve GPU utilization. Key aspects of the study include the design of dynamic computational graphs, pruning and quantization of neural networks, batch scheduling techniques, and integration of AI acceleration hardware. Furthermore, the article explores how SDVs create flexible platforms for deploying perception pipelines as scalable, modular services that respond adaptively to computational demands. Experimental evaluations highlight performance trade-offs between energy consumption, throughput, and latency reduction, offering insights into the deployment of perception models at scale. The findings demonstrate that a combination of model optimization and hardware-software co-design is critical for enabling safe, reliable, and efficient autonomous navigation. By aligning algorithmic efficiency with real-time constraints, autonomous vehicles can achieve high-precision perception while minimizing computational overhead. This article contributes a structured framework for advancing perception workflows through deep learning optimization in smart vehicles, with implications for the next generation of mobility services.