SCR-YOLO: An Efficient UAV Object Detection Framework Based on Spatial-Channel Reconstruction and Statistical Metric Learning

View at publisher Summarise with VIRAT
Authors: Xiuquan Yue, Tianshi Zhang, Hanyu Deng, Shiwei Wang, Junying Yuan, Haishan Chen, Haoran Deng
Cited by: 0
References: 10

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Unmanned Aerial Vehicle (UAV) imagery has emerged as a pivotal data source for intelligent social governance applications, such as urban traffic monitoring and crowd analysis. However, object detection in high-resolution aerial views faces severe challenges, particularly for tiny targets (e.g., pedestrians and vehicles). Most detectors suffer from computational redundancy in irrelevant background regions and gradient instability when localizing discrete tiny objects, primarily due to extreme background noise and the limitations of traditional geometric metrics in handling minute scale variations. To address these methodological and mathematical challenges, this paper proposes SCR-YOLO, an efficient detection framework based on structural optimization of YOLOv8. First, we design a novel Spatial-Channel Reconstruction C2f (SCR-C2f) module by integrating SCConv to replace the second standard convolution in the bottleneck unit. By integrating a redundancy-reducing topology, SCR-C2f effectively suppresses substantial background noise and refines the features of sparse target entities. Second, we introduce a collaborative regression strategy integrating Normalized Gaussian Wasserstein Distance (NWD). By modeling bounding boxes as 2D Gaussian distributions, this statistical metric mitigates the sensitivity and instability issues inherent in traditional IoU-based geometric metrics when processing tiny objects. Experimental results on the challenging VisDrone2019 benchmark demonstrate that our method significantly improves mAP50 by 5.6% and mAP50–95 by 1.2%, reaching 30.9% and 16.8%, respectively,with a real-time inference speed of 115 FPS on an NVIDIA RTX A4000 GPU. Consequently, the proposed method holds significant potential for deployment in intelligent social governance systems and resource-constrained UAV edge networks.

Fields of Study & Subjects

Advanced Neural Network Applications Visual Attention and Saliency Detection Domain Adaptation and Few-Shot Learning

Publication details

Journal / source
International Conference on Artificial Intelligence and Social Network Systems (AISNS)
Pages
96–100
Publication date
9 January 2026
Language
EN