Implementation of an Edge-Computing Vision System on Reduced-Board Computers Embedded in UAVs for Intelligent Traffic Management

Bemposta Rosende, Sergio; Ghisler, Sergio; Fernández-Andrés, Javier; Sánchez-Soriano, Javier · 2023 · Crossref

DOI: 10.3390/drones7110682

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of enhancing autonomous vehicle navigation in complex urban environments, specifically intersections and roundabouts, by integrating computer vision with unmanned aerial vehicles (UAVs). The authors identify a research gap in effectively combining low-power edge computing hardware with optimized computer vision models to provide real-time traffic management data. The primary objective is to develop, evaluate, and compare various object detection models and reduced-board computers to determine the optimal configuration for UAV-based edge computing, balancing inference speed, accuracy, and energy efficiency. The experimental design involved selecting four reduced-board computers: Raspberry Pi 3B+, Raspberry Pi 4, Jetson Nano, and Google Coral. Four deep learning models were chosen for evaluation: YOLOv5, YOLOv8, DETR, and EfficientDetLite. The researchers constructed a custom dataset by combining two existing aerial image datasets, resulting in 30,544 images labeled for cars and motorcycles. To optimize training, they applied a subsampling strategy (one frame per second) and used RoboFlow for image augmentation, yielding a final training set of 3,033 images. Model training was conducted on a high-performance computing cluster equipped with NVIDIA RTX 3080Ti GPUs. The trained models were then deployed on the selected edge devices to measure inference frames per second (FPS), model metrics, and energy consumption. The results indicated that the combination of the YOLOv8 model with the Jetson Nano board best suited the specific use case, offering an optimal balance of performance and efficiency. Alternatively, the EfficientDetLite model paired with the Google Coral board demonstrated significantly higher inference speeds but with lower accuracy compared to the YOLOv8/Jetson Nano configuration. The study highlights that while cloud computing offers scalability, edge computing on UAVs provides critical advantages for traffic management, including low latency, enhanced security, and energy efficiency by reducing data transmission requirements. The significance of this work lies in providing a validated framework for deploying intelligent traffic management systems using UAVs. By identifying the most effective hardware-software combinations, the findings assist autonomous vehicle manufacturers, transportation authorities, and software developers in optimizing real-time decision-making processes. The study contributes to the field by demonstrating that edge computing on reduced-board computers is a viable and superior solution for real-time object detection in dynamic traffic scenarios, addressing the limitations of current autonomous navigation systems in complex urban settings.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms