AHCO-YOLO: An Algorithm-Hardware Co-Optimization Framework for Energy-Efficient and Real-Time Object Detection on Edge Devices

초록

Real-time object detection on edge devices operates under tight computational, memory, and power budgets. Prior work typically treats model compression and hardware acceleration independently, yielding suboptimal trade-offs among accuracy, latency, and energy. We present AHCO-YOLO, an algorithm-hardware co-optimization framework (AHCO) that unifies model design, quantization, design space exploration (DSE), and hardware implementation. This approach overcomes the limitations of isolated methods and delivers synergistic gains. We introduce a hardware-friendly, lightweight You Only Look Once (YOLO) model with batch normalization (BN)-preserving quantization method that reduces the model size while maintaining accuracy at low precision. In addition, we propose a layer-specific resource-latency-aware DSE (LSRLA-DSE) method that selects the optimal dataflow based on layer-wise features and searches hardware design parameters under latency and resource constraints. Furthermore, we propose a FIFO-based streaming architecture with layer-wise dynamic dataflows that maintains high processing element (PE) utilization while minimizing off-chip traffic. Moreover, we introduce a semantic partition and regrouping strategy (SPRG) that improves resource efficiency and throughput. Implemented on a Xilinx ZCU104 FPGA, AHCO-YOLO-T achieves 79.8 FPS at 64.8% mAP, delivering 41.9 FPS/W and 80.5 GOPS/W. Across comparisons with existing YOLO accelerators, AHCO-YOLO achieves state-of-the-art efficiency, demonstrating suitability for real-time, energy-efficient object detection on edge platforms.

키워드