Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (21): 194-204.DOI: 10.3778/j.issn.1002-8331.2403-0136

• Graphics and Image Processing • Previous Articles     Next Articles

Research on Optimization of UAV Traffic Small Target Image Detection Algorithm

XU Huizhi, GU Xunan   

  1. School of Civil Engineering and Transportation, Northeast Forestry University, Harbin 150040, China
  • Online:2024-11-01 Published:2024-10-25

无人机视角下交通小目标图像检测算法优化研究

徐慧智,古旭楠   

  1. 东北林业大学 土木与交通学院,哈尔滨 150040

Abstract: Aiming at the small target detection algorithm of UAV images with low accuracy and prone to miss and false detection, an improved target detection algorithm based on YOLOv7 is proposed. A self-constructed pedestrian and vehicle dataset is utilized from UAV aerial video. The convolution in the neck and head is replaced with the CoordConv to improve the  ability to perceive the spatial information of the feature map. A small target detection layer is added to adapt to object targets at different scales and reduce the miss rate of the small targets, and the GE attention mechanism is added to the backbone network and the neck to enhance the utilization of contextual information. Wise-IoU is used as a bounding box loss function to introduce a dynamic non-monotonic focusing mechanism to improve the generalization ability of the model. The experimental results show that the improved algorithm has higher average precision than other algorithms with 91% accuracy in comparison experiments on self-built datasets. It is 2.1 percentage pionts higher than the YOLOv7 algorithm. Comparative experiments on the VisDrone2019 dataset show that the accuracy is 2.5 percentage pionts higher than that of the YOLOv7 algorithm, the comprehensive performance is better than other small-target detection algorithms, which verifies the generalization ability and effectiveness of the improved algorithm.

Key words: intelligent transportation, small target detection, deep learning, UAV images, YOLOv7 algorithm

摘要: 针对无人机图像的小目标检测算法精度低、容易出现漏检和误检的情况,提出一种基于YOLOv7的改进目标检测算法。利用无人机航拍视频自建行人和车辆数据集。将颈部和检测头中的卷积模块替换为CoordConv模块,提高算法感知特征图的空间信息能力;添加小目标检测层,适应不同尺度下的物体目标,降低小目标的漏检率;在主干网络和颈部增加GE注意力机制,加强上下文信息的利用。将Wise-IoU作为边界框损失函数,引入一种动态非单调聚焦机制,提高模型的泛化能力。实验结果表明,改进后的算法精度高于实验中其他算法,精度达到91%,比YOLOv7算法提升了2.1个百分点。在VisDrone2019数据集上进行对比实验,精度比YOLOv7算法提升了2.5个百分点;综合性能优于实验中其他小目标检测算法,验证了改进后算法的泛化能力与有效性。

关键词: 智能交通, 小目标检测, 深度学习, 无人机图像, YOLOv7算法