摘要
针对基于贝叶斯滤波器模型的目标跟踪算法非常依赖先验知识,在复杂场景中跟踪性能下降的问题,提出了一种神经过程和运动模型混合驱动的机动目标跟踪算法。利用神经网络构建一个目标运动的随机微分方程来提高对目标运动的建模能力;使用加速度模型作为领域知识约束目标状态的微分方程,构造一个混合驱动的解码器;利用所构成的混合驱动解码器替换神经过程的解码模块,形成一种无监督学习的混合驱动滤波器。仿真实验结果表明:所提算法同时具有数据驱动和模型驱动算法的优势,依赖先验知识少,能在不同场景下保持稳定的跟踪精度,生成的轨迹较神经过程滤波器更加平滑且具有目标动力学特征;与经典的贝叶斯滤波器相比,所提算法的状态估计峰值误差平均降低了20%。
Aiming at the problem that the target tracking algorithm based on Bayesian filter model relies heavily on prior knowledge and the tracking performance degrades in complex scenarios,a hybrid-driven maneuvering target tracking algorithm based on neural process and motion model was proposed.Firstly,a stochastic differential equation of target motion was constructed by using neural network to improve the modeling ability of target motion.Secondly,the acceleration model was used as the differential equation of the domain knowledge constraint target state to construct a hybrid-driven decoder.Finally,the decoding module of the neural process was replaced by the hybrid-driven decoder to form an unsupervised learning hybrid-driven filter.The simulation results show that the proposed algorithm has the advantages of both data-driven and model-driven algorithms.It relies less on prior knowledge and can maintain stable tracking accuracy in a variety of scenarios.The generated trajectory is smoother and has more target dynamic characteristics than the neural process filter.Compared with the classical Bayesian filter,the state estimation peak error of the proposed algorithm is reduced by 20% on average.
作者
朱洪峰
熊伟
崔亚奇
王子玲
ZHU Hongfeng;XIONG Wei;CUI Yaqi;WANG Ziling(Institute of Information Fusion,Naval Aviation University,Yantai,Shandong 264001,China)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2023年第4期152-161,共10页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金重大项目资助项目(61790554)
国家自然科学基金青年科学基金资助项目(62001499)。
关键词
机动目标跟踪
神经过程
混合驱动
无监督学习
状态估计
maneuvering target tracking
neural process
hybrid-driven
unsupervised learning
state estimation
