%0 Journal Article
%T 基于改进YOLOv5用于无人机图像的实时车辆检测<br>Real-Time Vehicle Detection Based on Improved YOLOv5 for Drone Images
%A 何佳桥
%A 葛晓东
%J Software Engineering and Applications
%P 705-713
%@ 2325-2278
%D 2023
%I Hans Publishing
%R 10.12677/SEA.2023.125068
%X 如今，无人机(UAV)逐渐被用于各个领域，例如交通监控和智能停车，其中对车辆的实时监测和分类是关键任务之一。在车辆检测方面存在许多挑战，例如小型车辆目标和无人机工作时飞行角度的变化导致目标尺度变化从而给车辆检测网络模型的优化带来的负担。此外，由于高空飞行时航拍图像目标较小、可提取特征较少，而导致模型检测精度较低。为解决上述问题，本文旨在基于YOLOv5算法，提出了一种准确、高效、实时的车辆检测网络。首先，为了使模型更好地提取小目标特征，在Neck部分添加了一个新的连接层，将第一个C3层的高分辨率特征映射连接到Neck部分。其次，为使模型更加专注小目标，添加一个步幅为4的输出层作为新的头部。同时优化模型对较大车辆目标的检测较少，我们去除掉了Head中输出特征图为20 × 20的检测头。同时，考虑到模型的推理速度，将Neck部分的C3模块替换为更轻量化的DS_C3模块。最后，为了进一步提高基于IOU损失函数的性能，将CIOU替换为α-IOU。本文使用VisDrone2019数据集，并基于改进算法和原始算法分别进行了实验，结果表明，本文的算法能够对对小目标进行有效的实时检测。<br />
Today, unmanned aerial vehicles (UAVs) are gradually being used in various fields, such as traffic monitoring and smart parking, where real-time monitoring and classification of vehicles is one of the key tasks. There are many challenges in vehicle detection, such as changes in flight angle when small vehicle targets and drones are working, resulting in changes in target scale, which puts a burden on the optimization of vehicle detection network models. In addition, due to the small aerial image target and fewer extractable features during high-altitude flight, the model detection accuracy is low. To solve the above problems, this paper aims to propose an accurate, efficient and real-time vehicle detection network based on the YOLOv5 algorithm. First, in order to make the model better extract small target features, a new connection layer is added to the Neck part, connecting the high-resolution feature mapping of the first C3 layer to the Neck section. Second, to make the model more focused on small targets, an output layer with a stride length of 4 is added as a new head. At the same time, the optimization model has less detection of larger vehicle targets, and we remove the detection head with an output feature map of 20 × 20 in Head. At the same time, considering the inference speed of the model, the C3 module of the Neck part was replaced with a more lightweight DS_C3 module. Finally, to further improve the performance of the IOU-based loss function, the CIOU is replaced with a α-IOU. This paper uses the VisDrone2019 dataset and conducts experiments based on the improved algorithm and the original algorithm, and the results show that the algorithm in this paper can effectively detect small targets in real time.
%K 车辆检测，YOLOv5，DSConv，自适应损失函数<br>Vehicle Detection
%K YOLOv5
%K DSConv
%K Adaptive Loss Function
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=74170