Comparative Analysis of YOLO-Based Algorithms for Vehicle Detection in Aerial Imagery

Dustali, Amin; Hasanlou, Mahdi; Azimi, Seyed Majid · 2025 · Crossref

DOI: 10.5194/isprs-archives-xlviii-g-2025-411-2025

archive: archived pipeline: cataloged verified

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

Summary

This paper addresses the challenge of selecting optimal real-time object detection algorithms for vehicle detection in aerial imagery, a task critical for applications like traffic surveillance and autonomous systems. The authors focus on the You Only Look Once (YOLO) family of models, which are favored for their balance of speed and accuracy. Specifically, the study compares four recent lightweight versions—YOLO-v8-n, YOLO-v9-t, YOLO-v10-n, and YOLO-v11-n—to determine which offers the best performance trade-off between inference speed and detection accuracy in challenging environments characterized by varying illumination and weather conditions. The experimental design utilized the EAGLE dataset, comprising aerial images captured at altitudes between 300m and 3000m. The images were preprocessed by cropping them into 1024 × 1024px tiles and resizing them to 416 × 416px for uniformity. Data augmentation techniques, including random rotations and flipping, were applied to enhance model robustness. The dataset was split into 23,001 training tiles and 7,682 validation tiles. The four YOLO models were fine-tuned for 10 epochs using the AdamW optimizer with a learning rate of 0.001 and a batch size of 8, leveraging pre-trained weights from the MS-COCO dataset. Performance was evaluated on an Nvidia Jetson AGX Xavier embedded GPU board using metrics including inference time, Average Precision (AP), and F1-Score. The results revealed distinct performance characteristics for each model. YOLO-v10-n achieved the fastest inference time at 4.00 seconds, attributed to its NMS-free training and spatial-channel decoupled downsampling, making it ideal for low-latency applications. Conversely, YOLO-v11-n demonstrated the highest detection accuracy, achieving an AP of 0.63 and the highest F1-Score of 0.95. This superior performance is linked to its reduced parameter count (22% fewer than YOLO-v8m) and the use of synthetic data generation during training. YOLO-v9-t exhibited the slowest inference time (4.53 seconds) despite incorporating Programmable Gradient Information and GELAN modules. YOLO-v8-n showed solid performance with an AP of 0.63 but lagged behind YOLO-v11-n in precision metrics. The study also highlighted that pre-trained models without fine-tuning performed poorly on aerial imagery, underscoring the necessity of domain-specific adaptation. The significance of this work lies in providing clear guidelines for deploying YOLO models on edge computing devices. The findings suggest that YOLO-v10-n is the optimal choice for applications prioritizing real-time processing speed, such as dynamic traffic monitoring. In contrast, YOLO-v11-n is recommended for scenarios requiring high precision in complex or data-scarce environments, such as disaster management or infrastructure monitoring. The paper concludes that understanding these specific strengths allows practitioners to select the most appropriate algorithm based on whether speed or accuracy is the primary constraint, thereby enhancing the efficiency of real-time surveillance and monitoring systems.

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.

• dms validation