基于改进YOLOv9的大重叠度无人机低空遥感影像目标检测方法

摘要:

无人机在执行低空飞行任务时,为了提升影像采集的精度和冗余度,会采用重叠航拍的方式。这种重叠航拍的方式虽然可以便于在后续处理中建立完整的地理空间信息模型,但重叠度过高也会导致数据冗余,增加后续数据处理的复杂度。为此,提出基于改进YOLOv9的大重叠度无人机低空遥感影像目标检测方法。通过阴影补偿和影像增强保证遥感影像质量,计算遥感影像重叠度调整量,移动并拼接相邻的无人机低空遥感影像。采用背景差分的方式,分割拼接遥感影像中的前景目标区域。构建YOLOv9网络,调整网络的连接方式并引入注意力机制,优化骨干网络以更紧凑地表示特征,减少重叠数据的冗余度,实现YOLOv9网络的改进。将分割的前景影像区域输入到改进YOLOv9算法中,得出遥感影像的特征提取结果。根据检测目标结构与纹理特征,设定目标标准特征。计算提取影像特征与设定标准特征之间的匹配度,根据匹配度与设定阈值之间的关系,得出无人机低空遥感影像的目标检测结果。通过效果测试实验得出结论:与传统检测方法相比,优化设计方法的影像特征提取一致性系数更高,目标检测的成功系数取值更大、目标位置检测误差更小,即优化设计方法的目标检测效果更优。

Abstract:

When performing low altitude flight missions, drones will use overlapping aerial photography to improve the accuracy and redundancy of image acquisition. Although this overlapping aerial photography method can facilitate the establishment of a complete geographic spatial information model in subsequent processing, excessive overlap can also lead to data redundancy and increase the complexity of subsequent data processing. Therefore, a high overlap unmanned aerial vehicle low altitude remote sensing image target detection method based on improved YOLOv9 is proposed. By using shadow compensation and image enhancement to ensure the quality of remote sensing images, the overlap adjustment amount of remote sensing images is calculated, and adjacent low altitude remote sensing images of unmanned aerial vehicles are moved and concatenated. Using background subtraction to segment and concatenate foreground target areas in remote sensing images. Build YOLOv9 network, adjust the connection mode of the network and introduce attention mechanism, optimize the backbone network to represent features more compactly, reduce the redundancy of overlapping data, and achieve the improvement of YOLOv9 network. Input the segmented foreground image regions into the improved YOLOv9 algorithm to obtain the feature extraction results of remote sensing images. Set target standard features based on the detection target structure and texture features. Calculate the matching degree between the extracted image features and the set standard features, and based on the relationship between the matching degree and the set threshold, obtain the target detection result of the unmanned aerial vehicle low altitude remote sensing image. The conclusion drawn from the effect testing experiment is that compared with traditional detection methods, the optimized design method has a higher consistency coefficient for image feature extraction, a larger success coefficient for object detection, and a smaller error in object position detection, indicating that the optimized design method has a better object detection effect.