Résumé
Deep learning-based computer vision methods have significant potential for monitoring construction sites. However, deploying Convolutional Neural Networks (CNNs) on battery-powered edge devices presents significant computational and energy efficiency challenges. Few studies systematically explore trade-offs between accuracy, latency, throughput, and energy consumption. This paper presents an FPGA-based CNN inference methodology specifically for construction progress monitoring, evaluated on two distinct hardware platforms: a low-cost Xilinx Zynq-7020 and a high-performance Xilinx UltraScale+. By leveraging quantization methods, including Quantization-Aware Training (QAT) with QKeras and hardware implementation via the hls4ml framework, the accelerators achieved over 500× latency reduction and approximately 180× energy efficiency compared to ARM CPU inference, with accuracy exceeding 90%. These results demonstrate the practical viability of deploying CNNs in real-time, energy-constrained environments.