Edge-Optimized YOLO Architectures for Real-Time Autonomous Vehicle Perception: A Hardware-Aware Co-Design Framework

Abstract

Object detection is the perceptual backbone of autonomous driving, and single-stage detectors of the YOLO family have become the practical default wherever frames must be processed within real-time budgets. Yet the published detection literature optimizes predominantly for benchmark accuracy on server-class accelerators, while vehicles impose a different objective: bounded end-to-end latency at high frame rates, on power- and thermally-constrained embedded accelerators, across multiple simultaneous camera streams, under safety expectations that penalize missed detections far more than the mean average precision metric reflects. This paper develops a research concept for an edge-optimized YOLO-based perception component designed and evaluated against vehicle-grade constraints. The concept specifies a co-design space spanning architecture, compact backbones, decoupled heads, resolution and anchor policy tuned to driving object statistics, and compression, structured pruning, quantization-aware training to integer arithmetic, and knowledge distillation from a high-capacity teacher, searched jointly under hardware-in-the-loop latency measurement rather than proxy operation counts. A deployment architecture allocates per-camera detection instances across embedded accelerator resources with a frame-freshness scheduling policy that privileges recency over throughput, and an optional offload path is analyzed and deliberately excluded from the safety path. The evaluation plan is phased: accuracy and robustness on driving benchmarks with corruption suites; latency, jitter, energy, and thermal behavior measured on target hardware across compression configurations; and system-level metrics that couple detection quality to reaction distance at speed. The concept's central claim is methodological: for vehicle perception, the deployable operating point is a property of the model-compression-hardware triple, and it must be measured as such.