YOLOX-IM:一种无人机航拍视频的轻量化交通参数提取模型

YOLOX-IM: A lightweight traffic parameter extraction model for UAV aerial images

在无人机航拍的过程中,背景更广阔,目标的尺寸更小,种类更复杂。提出一种基于YOLOX-s的轻量化无人机航拍目标检测算法YOLOX-IM。首先,为了提高该模型检测小目标的性能,通过使用切片辅助推理(slicing aided hyper inference,SAHI)算法以及坐标修正矩阵对训练集进行预处理和数据增强。然后,在路径聚合网络(path aggregation network,PAN)中引入一个浅层特征图以及超轻量级子空间注意模块,并添加一个检测头对小物体进行检测;最后,对边界回归的损失函数进行了优化。在VisDrone2019数据集的消融实验结果表明,所提出的模型检测精度与基础YOLOX-s相比高了8.13%;模型体积4.55 MB,相较于原模型下降67.14%。利用该模型在中国天津市渌水道进行实地交通监测的交通参数提取实验,在设定的场景中,当无人机航拍高度为50 m时,该模型的交通提取参数精度最高,达到96.14%。

In the process of UAV aerial photography,the background is broader and the targets are smaller in size.In this paper,we propose a lightweight target detection algorithm YOLOX-IM for UAV aerial photography based on YOLOX-s.First,to improve the performance of small target decection,the training set is preprocessed and data is enhanced by using a slicing aided hyper inference(SAHI)algorithm as well as a coordinate correction matrix.Then,a shallow feature map as well as an ultra-lightweight subspace attention module are introduced in the path aggregation network(PAN),and a detection head is added for small object detection.Finally,the loss function of the boundary regression is optimized.The experimental results on the VisDrone2019 dataset show that the proposed model has 8.13%higher detection accuracy compared with the YOLOX-s;compared with the original model,the model volume is significantly reduced to 4.55 MB,which is 67.14%lower than the original model.Next,the model is used to conduct traffic parameter extraction for field traffic monitoring in Tianjin,Lushui road,China.The study indicates that the model has the highest traffic extraction parameter accuracy of 96.14%at the UAV altitude of 50 m.