基于航拍图像的自适应感知目标检测网络  

Adaptive perception object detection network based on aerial photography

作　　者：袁玲玲 陈春梅[1,2] 朱天鑫 邓豪 刘桂华[1,2] Yuan Lingling;Chen Chunmei;Zhu Tianxin;Deng Hao;Liu Guihua(School of Information Engineering,Southwest University of Science and Technology,Mianyang 621000,China;Robot Technology Used for Special Environment Key Laboratory of Sichuan Province,Mianyang 621000,China)

机构地区：[1]西南科技大学信息工程学院,绵阳621000 [2]特殊环境机器人技术四川省重点实验室,绵阳621000

出　　处：《电子测量技术》2025年第2期57-65,共9页Electronic Measurement Technology

基　　金：中央军委装备发展部项目(23ZG8102);校博士基金(20zx7123)项目资助。

摘　　要：由于无人机拍摄高度和角度的多样性,其图像往往呈现背景复杂且小目标居多的特征,这导致了相关检测算法性能较差。针对此问题,本文提出了一种基于自适应感知网络的航拍图像车辆检测方法,旨在从提高车辆特征显著度和改善特征信息损失两个方面来提升小目标的检测性能。首先,为了提取更高效的特征表征,提出了自适应感知特征提取模块,该模块通过捕捉长程依赖关系和更强的几何特征表示,能够自适应地对物体的形状进行建模。其次,为了减少下采样和连续池化造成的信息损失,设计了双分支空间感知下采样模块,该模块混合不同通道的特征图,以最大限度地保留小目标特征信息。然后,在特征融合网络中,引入了具有丰富空间信息的浅层特征图,以增强小目标的检测能力。最后,设计了新的动态回归损失函数DEIoU,该函数引入惩罚项来度量真实框与检测框之间横纵比的相关性,从而进一步提高网络的预测精度。在Visdrone数据集上的实验结果表明,所提方法的平均精度均值mAP达到了70%,推理速度达到了99.26 fps,实现了较好的速度与精度的平衡,并且所提方法在UCAS-AOD数据集上取得了最佳的检测精度,具有较强的泛化能力。Due to the diverse height and angle of drone shots,the images often have complex backgrounds and mainly feature small targets.As a result,the performance of detection algorithms for these images is often poor.To address this issue,this paper presents a vehicle detection method for aerial images using an adaptive perception network.The goal is to improve the detection of small targets by focusing on two aspects:enhancing the saliency of vehicle features and improving the preservation of feature information.First,an adaptive perception feature extraction module is proposed to extract a more efficient feature representation.This module captures long-range dependencies and stronger geometric feature representations to adaptively model the shape of objects.Second,a dual-branch spatial perception downsampling module is introduced to mitigate information loss caused by down-sampling and continuous pooling.This module combines feature maps of different channels to maximize the retention of small target feature information.Next,the feature fusion network incorporates shallow feature maps with rich spatial information and adds detection heads to enhance the detection capability of small targets.Finally,a new dynamic regression loss function,DEloU,is designed.This function includes a penalty term to measure the correlation between the aspect ratio of the ground truth box and the detection box,further improving the prediction accuracy of the network.Experimental results on the Visdrone dataset show that the proposed method achieves an average precision(mAP)of 69.9%and an inference speed of 99.26 fps,indicating a good balance between speed and accuracy.Moreover,the proposed method has achieved the best detection accuracy on the UCAS-AOD dataset and has strong generalization ability.

关 键 词：无人机 目标检测 自适应感知特征提取 特征融合网络 双分支空间感知下采样 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN06[自动化与计算机技术—计算机科学与技术]