计算机应用

• 人工智能与仿真 •    下一篇

基于改进PP-YOLO和Deep-SORT的多无人机实时跟踪算法

马峻,姚震,徐翠锋,陈寿宏   

  1. 桂林电子科技大学电子工程与自动化学院
  • 收稿日期:2021-07-02 修回日期:2021-09-14 发布日期:2021-09-24 出版日期:2021-09-24
  • 通讯作者: 姚震

Multiple drones real-time tracking algorithm based on improved PP-YOLO and Deep-SORT

  • Received:2021-07-02 Revised:2021-09-14 Online:2021-09-24 Published:2021-09-24

摘要: 无人机(UAV)目标尺寸较小,多架无人机之间特征也不明显,鸟类和飞虫的干扰给无人机目标的准确检测和稳定跟踪带来了巨大挑战。针对传统目标检测算法对小目标无人机检测性能差、跟踪不稳定的问题,提出一种基于改进PP-YOLO和Deep-SORT的多无人机实时跟踪算法。首先,将压缩-激励模块融入到PP-YOLO检测算法中,实现对无人机目标的特征提取和检测;其次在ResNet50-vd结构中引入Mish激活函数,消除反向传播过程中的梯度消失问题,进一步提升检测精度。然后,采用Deep-SORT算法实时跟踪无人机目标,将提取外观特征的主干网络更换为ResNet50,改善原有网络对微小外观感知能力弱的问题;最后引入损失函数Margin Loss,既提高了类别可分性,又加强了类内紧度和类间差异。实验结果表明,平均检测精度相比原始PP-YOLO算法提升了2.27%,跟踪准确性相对于原始Deep-SORT算法提升了4.5%。改进后的算法跟踪精度可达91.6%,能够实时跟踪600米以内多架无人机目标,有效解决了跟踪过程中的“丢帧”问题。

Abstract: The target size of the Unmanned Aerial Vehicle (UAV) is small, and the characteristics among multiple UAVs are not obvious. The interference of birds and flying insects brings a huge challenge to the accurate detection and stable tracking of the UAV target. Aiming at the problem of poor detection performance and unstable tracking of small target UAVs by traditional target detection algorithms, a real-time tracking algorithm for multiple UAVs based on improved PaddlePaddle-YOLO (PP-YOLO) and Simple Online and Realtime Tracking with a Deep association metric (Deep-SORT) was proposed. First, the squeeze-excitation module was integrated into the PP-YOLO detection algorithm to achieve feature extraction and detection of UAV targets; Secondly, the Mish activation function was introduced into the ResNet50-vd structure to eliminate the problem of gradient disappearance in the back propagation process and further improved the detection accuracy. Then, the Deep-SORT algorithm was used to track UAV targets in real-time, and the backbone network that extracts appearance features was replaced with ResNet50 to improve the original network’s weak perception of small appearances. Finally, the loss function Margin Loss was introduced, which not only improved the class separability, but also strengthened the tightness within the class and the difference between classes. The experimental results illustrate that the average detection accuracy is increased by 2.27% compared to the original PP-YOLO algorithm, and the tracking accuracy is increased by 4.5% compared to the original Deep-SORT algorithm. The improved algorithm has a tracking accuracy of 91.6% and can track multiple UAV targets within 600 meters in real-time, and effectively solving the problem of "frame loss" in the tracking process.

中图分类号: