A Method of Multi-Motion Target Recognition Based on Deep Learning for UAV Platforms
LI Xun
ZHANG Qianfei
LIU Xin
CHEN Weizhong
ZHOU Huilong
Abstract：In the light of the limited problems in model scalability and precision that scale is large in varia-tion,dense is irregular in distribution,and class is imbalance in vehicle object detection on drones,this study proposes an improved YOLO-QYF vehicle detection method based on the YOLOv7.This method is to introduce the QARepVGG module to replace the computationally intensive E-ELAN in the baseline,and minimize information loss during feature map transmission,a content aware reassembly of features is em-ployed,and a coordinate attention mechanism is integrated to enhance the localization ability for objects of interest.Additionally,the WIoU loss function is introduced to address the class imbalance issue and im-prove the model's generalization capability.The experimental results demonstrate that the proposed method achieves the mean average precisions of 96.4％,95.6％and 94.5％under the free flow,the syn-chronous flow and the blocking flow traffic scenarios respectively,increasing by 2.0％,1.6％ and 3.4％compared to the baseline,and the frames per second are 81.62,78.13 and 76.34 respectively.In the vis-drone2021 dataset,the mean average precisions of the proposed method achieves 60.6％,and is 1.4％higher than the baseline.
Keywords：UAV vehicle recognitionYOLOv7QARepVGGCARAFEcoordinate attention
Publication Date：2025-12-25
Online Publishing Date：2025-12-26(First online date of this platform, not the publication date of the document)
Pages：11( 85-95 )