YOLOv8-PSA: A YOLOv8 Enhanced by Partial Self-Attention Mechanism for Urban UAV Detection

Abstract

Unmanned Aerial Vehicle (UAV) detection in urban environments is critical for airspace management and emergency response. However, dynamic scale variations, complex backgrounds, and real-time requirements hinder existing methods. This research proposes YOLOv8-PSA, which integrates a Partial Self-Attention (PSA) mechanism to capture global contextual dependencies and suppress background noise. A multi-scale feature fusion module with a lightweight C2fCIB block enhances detection accuracy across varying scales. CIoU loss is adopted to optimize bounding box localization. Experiments on an urban UAV dataset demonstrate 97.8% precision, 92.5% recall, and 96.5% mAP50, with consistent improvements over the baseline YOLOv8n. Comparative and ablation studies validate the effectiveness of the proposed modules for practical urban UAV detection.