Resumen
In the 5G intelligent edge scenario, more and more accelerator-based single-board computers (SBCs) with low power consumption and high performance are being used as edge devices to run the inferencing part of the artificial intelligence (AI) model to deploy intelligent applications. In this paper, we investigate the inference workflow and performance of the You Only Look Once (YOLO) network, which is the most popular object detection model, in three different accelerator-based SBCs, which are NVIDIA Jetson Nano, NVIDIA Jetson Xavier NX and Raspberry Pi 4B (RPi) with Intel Neural Compute Stick2 (NCS2). Different video contents with different input resize windows are detected and benchmarked by using four different versions of the YOLO model across the above three SBCs. By comparing the inference performance of the three SBCs, the performance of RPi + NCS2 is more friendly to lightweight models. For example, the FPS of detected videos from RPi + NCS2 running YOLOv3-tiny is 7.6 times higher than that of YOLOv3. However, in terms of detection accuracy, we found that in the process of realizing edge intelligence, how to better adapt a AI model to run on RPi + NCS2 is much more complex than the process of Jetson devices. The analysis results indicate that Jetson Nano is a trade-off SBCs in terms of performance and cost; it achieves up to 15 FPSs of detected videos when running YOLOv4-tiny, and this result can be further increased by using TensorRT.