基于改进YOLOv7-tiny的无人机视角目标检测算法

UAV Perspective Object Detection Algorithm Based on Improved YOLOv7-tiny

针对无人机在进行目标检测任务时存在背景复杂、小目标多且密集以及无人机的硬件条件有限等问题,基于YOLOv7-tiny,提出一种改进的无人机视角目标检测算法。首先,将原主干网络结合MobileNetV3进行轻量化改进,引入深度可分离卷积结构,减少网络的参数量;其次,在颈部网络添加GAM全局注意力机制,使网络能够聚焦目标特征,提升检测性能;最后,将定位损失函数改进为SIoU函数,提高网络训练的收敛速度及精度。在VisDrone2019数据集上进行的消融实验及对比实验显示,本文改进后的算法平均检测精度达到45.5%,mAP50达到37.6%,浮点运算量GFLOPS为10.2 G/S,与原算法相比,精度提高了1.7%,mAP50提高了4.2%,GFLOPS减小23%,每一个改进模块都有效提升了无人机检测复杂背景中小目标的能力。

Aiming at the problems of complex background,multiple and dense small targets,and limited hardware conditions of drones in target detection tasks,an improved drone perspective target detection algorithm based on YOLOv7-tiny is proposed.Firstly,the original backbone network is combined with MobileNetV3 for lightweight enhancement,introducing a deep separable convolutional structure to reduce the number of network parameters;Secondly,a Global Attention Mechanism(GAM)is added to the neck network to enable the network to focus on target features and improve detection performance;Finally,the localization loss function is improved to the SIoU function to improve the convergence speed and accuracy of network training.The ablation and comparative experiments conducted on the VisDrone2019 data set showed that the improved algorithm in this paper achieved an average detection accuracy of 45.5%,mAP50 of 37.6%,and a floating-point operation GFLOPS of 10.2 G/S.Compared with the original algorithm,the accuracy was improved by 1.7%,mAP50 of 4.2%,and GFLOPS of 23%.Each improved module can effectively enhance the ability of drones to detect small targets in complex backgrounds.