YOLOv5-NCA: drone detection model with multiscale framework and feature fusion

Abstract

Detecting drones is crucial for countering unauthorized flights. However, real-world scenarios with complex backgrounds and small targets frequently cause false positives and missed detections. To address this challenge, this paper proposes YOLOv5-NCA, which builds on YOLOv5s and adopts the Normalized Wasserstein Distance (NWD) loss to boost smalltarget detection accuracy. An attentional scale sequence fusion (ASF) framework is applied to strengthen multi-scale feature extraction. A context aggregation attention (CAA) module is inserted to enrich feature information and improve detection in complex backgrounds. The C3 module is upgraded to the C3_F module, reducing computational cost, parameter volume, and deployment difficulty. The designed YOLOv5-NCA achieves 84.6% average precision on the processed anti-UAV dataset, a 3.7% improvement over YOLOv5s. Compared with mainstream one-stage detectors, YOLOv5-NCA delivers superior performance with a more compact model size.