A STUDY OF YOLO NEURAL NETWORK PERFORMANCE FOR EMBEDDED SYSTEMS AND UAVS
DOI: 10.31673/2412-4338.2025.038720
Abstract
The article presents a comprehensive study of the performance of modern YOLO (You Only Look Once) neural network versions across a wide range of hardware platforms — from desktop GPUs to mobile ARM devices and specialized AI accelerators. The study evaluates the efficiency of YOLOv4, YOLOv5, YOLOv8, as well as the latest YOLOv10n and YOLOv11n models, taking into account architectural complexity, optimization techniques (FP16, INT8, TensorRT), inference latency, power consumption, and frames per second (FPS). To ensure objective cross-platform comparison, a unified performance unit system was proposed, based on the reference execution of YOLOv4 on an NVIDIA GTX 1070 GPU. More than 1000 runs were conducted on real hardware, including Orange Pi 5/5 Plus, Raspberry Pi 5
(with Hailo-8L AI HAT), Rockchip RK35xx series, Google Coral platforms, and the NVIDIA Jetson Orin lineup (Nano, NX, AGX). All measurements were performed with statistical significance (50–100 iterations) and a confidence interval of 90–95 %. Additionally, inference modes, quantization types, and hardware integration aspects were considered. The results identified optimal configurations for various real-time application scenarios. Special attention was given to YOLO deployment onboard drones, where limited energy resources, device weight, compactness, autonomy, and low latency are critical. Typical embedded architectures were analyzed, and the deployment efficiency of models under platform-specific constraints was evaluated. The article provides generalized recommendations for selecting models and hardware solutions for autonomous navigation, monitoring, surveillance, and robotics systems operating under strict power budgets. The findings can be valuable to developers and researchers in the fields of embedded systems, Edge AI, IoT, and autonomous unmanned platforms.
Keywords: YOLO, deep learning, embedded systems, drones, inference, energy efficiency, GPU, ARM, Google Coral, NVIDIA Jetson Orin
