基于无人机平台的多目标跟踪算法  

Multi-target Tracking Algorithm Based on UAV Platform

作　　者：任进[1] 李文邦 郭昱汝 REN Jin;LI Wenbang;GUO Yuru(School of Information Science and Technology,North China University of Technology,Beijing 100144,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《无线电工程》2023年第1期34-39,共6页Radio Engineering

基　　金：北京市优秀人才培养资助青年骨干个人项目(401053712002);2022年北京市大学生创新创业训练计划项目(22XN238)。

摘　　要：针对目前无人机平台多目标跟踪技术的跟踪精确度低、占用内存大的问题,提出了一种基于不同检测器算法和DeepSort算法结合而成的多目标跟踪算法,提高在无人机上对地面行人在跟踪数据集中的效果。使用深度学习的多目标跟踪技术通过构建卷积神经网络(Convolutional Neural Network, CNN),用卡尔曼滤波算法实现了对目标轨迹的预测,匈牙利算法则使卡尔曼滤波的预测结果得以分配,使DeepSort算法在保证跟踪效果的同时,也保证了跟踪时的速度。实验结果显示,DeepSort在与YOLOv5x检测器配合后,多目标跟踪精度可提高20%。To address the issues of low tracking accuracy and large memory occupation of current multi-target tracking technology on UAV platform, a multi-target tracking algorithm based on different detector algorithms and DeepSort algorithm is proposed to improve the effect of tracking data set for pedestrians on the ground on UAV. The multi-target tracking technology using deep learning realizes the prediction of target trajectory with Kalman filter algorithm by constructing a Convolutional Neural Network(CNN). The prediction results of Kalman filter is distributed by the Hungarian algorithm, so the DeepSort algorithm ensures not only the tracking effect, but also the tracking speed. The experimental results show that the multi-target tracking accuracy can be improved by 20% by the combination of DeepSort and YOLOv5x detector.

关 键 词：无人机平台 多目标跟踪 目标检测 YOLOv5 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]