Implementing a YOLOv5 neural network in systems-on-chips: A comparative study between the NVIDIA SoC and the APSoC-Zynq7000

The aim of this article is to make a comparison between two high-performance maps used to implement convolutional neural networks such as YOLO. We explore two distinct approaches to this comparison. In the first approach, we implement directly on the Pynq-Z1 board by performing several trials and solving various problems that arise. The second approach involves implementing the network on the Jetson Nano board. This method involves first training the model on a main computer and then running it on the embedded system.Integrating the YOLO algorithm into the SoC presents opportunities to improve the perception and processing capabilities of embedded systems. However, it is important to note that this integration is not without its difficulties and limitations. In this article, we examine in detail the benefits, difficulties and limitations inherent in this integration. We focus our analysis on two separate development boards: the Pynq-Z1, which is based on an FPGA architecture, and the Jetson Nano, which is equipped with an NVIDIA SoC.Our ultimate goal is to achieve a comprehensive comparison between the performance of these two boards. We aim to identify the optimal results and determine which approach offers maximum efficiency. In summary, this paper provides an in-depth look at the positive aspects and challenges of integrating the YOLO algorithm into embedded systems using two distinct development platforms.