Benchmark Analysis of YOLO Performance on Edge Intelligence Devices

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