基于PHD图和序列的深度数据关联算法  

Deep Data Association Algorithm Based on PHD Graphs and Sequences

作　　者：刘龙 谢家强 张梦璇 胥庆 刘希龙 LIU Long;XIE Jiaqiang;ZHANG Mengxuan;XU Qing;LIU Xilong(School of Electronic Engineering,Xidian University,Xi’an,Shaanxi 710071,China;School of Artificial Intelligence,Xidian University,Xi’an,Shaanxi 710071,China)

机构地区：[1]西安电子科技大学电子工程学院,陕西西安710071 [2]西安电子科技大学人工智能学院,陕西西安710071

出　　处：《信号处理》2024年第11期2062-2073,共12页Journal of Signal Processing

基　　金：陕西省技术创新引导计划(2023KXJ-279)。

摘　　要：随着跟踪环境的日益复杂,杂波、目标密集容易造成关联错误,且最优匹配过程随目标数的增长排列组合也呈指数增长,多目标航迹关联的问题愈加凸显,增加了传统算法的复杂性和不可靠性。针对目标数变化和杂波干扰造成的最优匹配问题,本文提出了一种基于概率假设密度(Probability Hypothesis Density,PHD)图特征和序列特征的深度关联网络,首先将目标的PHD图与量测的PHD图进行融合以挖掘目标与量测空间关联特征以及空间结构信息,进而将点迹特征扩展为图像特征,增加可获取的关联判别依据。然后本文对点迹特征或序列特征依据深度学习构建新的度量方法,减少度量方法对于距离的依赖性,从而削弱密集杂波造成的误判。最后,为了避免目标数和量测增加时导致的组合爆炸,本文提出一种基于多头注意力机制的深度匹配网络,利用8头注意力机制关注邻域的分配情况以提高匹配过程的准确性和可靠性,同时可减少计算耗时。消融实验证明,图像特征、序列特征、深度匹配网络均可提高关联信息的特征提取能力以及全局最优匹配过程。仿真实验证明,该深度关联网络可以从隶属矩阵和最小指派过程同时提升关联的准确性和稳定性。此外,该方法也说明了将现有的图像跟踪的关联算法迁移到雷达跟踪中的可行性,为雷达跟踪数据关联的发展提供新思路。Clutter and dense targets often cause association errors as tracking environments become increasingly complex.The optimal matching process grows exponentially with the number of targets,highlighting the complexity and unreliability of traditional algorithms for multi-target track association.This study proposes a deep association network based on probability hypothesis density(PHD)map features and sequence features to address optimal matching issues caused by varying target numbers and clutter interference.First,the PHD map of targets was fused with the PHD map of measurements to explore spatial association features and structural information.Then,point features were extended to image features,increasing the basis for association discrimination.The study developed a new measurement method based on deep learning for point or sequence features,reducing the dependency on distance and thereby mitigating misjudgments caused by dense clutter.Finally,a deep matching network based on a multi-head attention mechanism was proposed to avoid combinatorial explosion as the number of targets and measurements increases.This network used an eight-head attention mechanism to focus on neighborhood distribution,improving the accuracy and reliability of the matching process while reducing computational time.Ablation experiments demonstrated that image features,sequence features,and the deep matching network enhance the feature extraction capabilities for association information and the global optimal matching process.Simulation experiments showed that this deep association network improves the accuracy and stability of association in both the affinity matrix and the minimum assignment process.Additionally,this method illustrates the feasibility of transferring existing image tracking association algorithms to radar tracking,providing new insights for developing radar tracking data association.

关 键 词：数据关联 概率假设密度图 隶属度矩阵 空间相似度 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]