结合金字塔池化和运动轨迹的目标跟踪方法  

Single Target Tracking Algorithm Based on Pyramid Pooling and Target Motion Trajectory

作　　者：崔伟 段迅[1] 孔广黔[1] 龙慧云[1] CUI Wei;DUAN Xun;KONG Guang-qian;LONG Hui-yun(School of Computer Science and Technology,Guizhou University,Guiyang Guizhou 550025,China)

机构地区：[1]贵州大学计算机科学与技术学院,贵州贵阳550025

出　　处：《计算机仿真》2025年第3期216-225,共10页Computer Simulation

基　　金：国家自然科学基金(62266011);贵州省基础研究计划项目(黔科合基础-ZK[2022]一般119)。

摘　　要：针对现有目标跟踪算法存在目标感知能力较弱、实时性不足以及目标易丢失的问题,提出一种结合金字塔池化和目标运动轨迹的单目标跟踪算法(Pyramid pooling Transformer for single object tracking,PPTTrack)。首先,算法采用PVT作为骨干网络构建深度神经网络,建立模板图和搜索图之间的双向信息流通道,提升跟踪算法的目标感知能力;其次,在骨干网络中引入金字塔池化方法缩短K,V的序列长度,降低模型复杂度,提高算法运行速率;最后,将跟踪目标的运动轨迹构建为一个Motion Token输入到编码器中进行特征融合,利用运动轨迹连续性预测目标位置,以解决目标丢失问题。实验结果表明,上述算法在GOT-10K、LaSOT、UAV123以及TrackingNet四个基准数据集上的性能均达到了较为先进的水平。Aiming at the problems of weak target perception ability,insufficient real-time performance,and easy loss of targets in existing target tracking algorithms,a Pyramid Pooling Transformer for Single Object Tracking(PPTTrack)is proposed,which combines pyramid pooling and target motion trajectory.Firstly,the algorithm uses PVT as the backbone network to build a deep neural network,which establishes a bidirectional information flow channel between the template graph and the search graph to improve the target perception capability of the tracking algorithm.Token is input to the encoder for feature fusion,and the motion trajectory continuity is used to predict the target position in order to solve the target loss problem.The experimental results show that the algorithm achieves a more advanced performance on four benchmark datasets,namely GOT-10K,LaSOT,UAV123 and TrackingNet.

关 键 词：单目标跟踪 注意力机制 金字塔池化 运动轨迹 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]