关键词: YOLOv8, 无人机影像, 车辆检测, 注意力机制, 特征融合

Abstract: Accurate and real-time vehicle detection and tracking provide crucial data support for traffic flow estimation and intelligent traffic management.Drone imagery has emerged as a vital data source for vehicle detection tasks.To address the weak ability of existing YOLO models to detect small objects within complex scenarios and the scarcity of open-source datasets for drone vehicle detection,this paper proposes the YOLOv8s-VOD model specifically designed for vehicle detection tasks,and introduces the open-source dataset NJVOD.This method constructs C2f-PTB and BiFPN-GLSA modules to achieve collaborative extraction of global-local featuresand effective fusion of multi-scale semantic and edge information,thereby improving detection accuracywhile reducing network complexity.Experimental results show that YOLOv8s-VOD achieves the highest detection accuracy with minimal parameters,outperforming existing methods by 2.4～12.2 percentage poin on the VEDAI dataset and 4.1～5.3 percentage point on the NJVOD dataset;The C2f-PTB and BiFPN-GLSA modules proposed in this work both effectively improve small-object detection accuracy.Additionally,the newly created NJVOD dataset offers crucial support for related research.

Key words: YOLOv8, drone imagery, vehicle detection, attention mechanism, feature fusion