Edge-Oriented Evaluation of CNN-Based Vehicle Detection on Embedded V2X Platforms

Vehicle-to-Everything (V2X) systems increasingly depend on low-latency perception pipelines that must operate efficiently across the cloud-edge-vehicle continuum, often under constrained embedded hardware. While Convolutional Neural Networks (CNNs) achieve strong accuracy on high-end platforms, their performance shifts significantly in edge environments due to limited CPU parallelism, memory bandwidth bottlenecks, cache pressure, and degraded sensor input quality. Building on our earlier offline evaluation of state-of-the-art detectors, this study investigates the real-time deployment feasibility of representative single-stage and two-stage CNN architectures, specifically for vehicle detection, on embedded$\mathbf{V} 2 \mathbf{X}$compute nodes, including near-edge Roadside Units (RSUs) and far-edge in-vehicle platforms. Empirical results reveal a pronounced divergence: lightweight single-stage models sustain real-time performance by leveraging compact activation flows and shallow feature hierarchies, while region-based and deeper networks incur severe latency overheads due to memory-bound operations and constrained instruction-level parallelism. Our deployment-based measurements underscore the limits of complex perception models at the edge and emphasize the need for a re-architected V2X perception stack that balances model complexity with hardware constraints, sensor fidelity, and cooperative sensing capabilities.