Real-time Object Detection on FPGA-based Heterogeneous MPSoCs: A Preliminary Analysis of the Execution Bottlenecks

FPGA-based heterogeneous Multiprocessor Systems-on-Chips (MPSoCs) provide an efficient platform for real-time object detection by combining programmable logic with embedded processors. This paper evaluates the deployment of Tiny YOLOv2 on an AMD Xilinx Zynq UltraScale+ MPSoC, demonstrating a throughput of 30 FPS with a power consumption of just 1.1 W—significantly surpassing GPU (15 FPS, 13.5 W) and CPU (2.5–8 FPS, 2.1–5.3 W) implementations. Additionally, a preliminary analysis of execution bottlenecks using the Vitis AI Profiler is conducted, providing valuable insights into performance limitations and identifying opportunities for further optimization of real-time applications on similar platforms.