摘要
针对传统机动目标跟踪问题中存在的跟踪精确性低、参数适应性差、计算量大等问题,提出了一种基于径向基神经网络的多模型机动目标跟踪方法。首先,介绍了机动目标跟踪问题的基本原理、径向基神经网络模型及机动目标运动模型。然后,将提取的机动目标特征向量输入已训练好网络参数的神经网络中,与隐含层中由训练样本组成的输入矩阵比较并输出,通过融合多个观测模型的状态估计,得到统一的目标状态估计值,从而建立了基于神经网络的切换多模型算法原理架构,并给出相应的计算框图。最后,通过数学仿真,比较了两种神经网络模型在全观测与部分观测条件下的目标状态估计性能。仿真结果显示,在目标进行大机动时,广义回归神经网络的观测误差方差更优,而基于径向基网络的切换多模型方法的鲁棒性更佳,其性能可提高11%。另外,基于径向基网络切换多模型方法的相关参数容易训练且易于在轨实时计算,具有更广泛的应用前景。
For the low tracking precision, poor parameter adaptability and large calculation quantity in the traditional maneuvering target tracking, a multi-model maneuvering target tracking method based on radial-based neural network is proposed. First, the basic principles of maneuvering target tracking, radial-based neural network model and maneuvering target motion model are introduced in this paper. Then, the extracted mobile target feature vectors are input into the neural network with trained network parameters, and the compared results with the input matrix composed of training samples in the hidden layer are output. By integrating the state estimation of multiple observation models, an unified target state estimate is given, and the principle architecture of switching multiple model algorithm based on the neural network and the corresponding calculation block diagram are established. Finally, the performance of target state estimation between two neural network models under full and partial observation are compared through mathematical simulation. Simulation results show that the observed error variance is better for a large target maneuver, while the switching multi-model method based on RBF network is more robust by 11%. In addition, the relevant parameters of the switching multi-model method based on RBF network are easy to train and to perform in-orbit real-time calculation,which has a wider application prospect.
作者
张晓杰
汪灏
赵灵峰
ZHANG Xiaojie;WANG Hao;ZHAO Lingfeng(Shanghai Engineering Center for Microsatellites,Shanghai 201210,China)
出处
《无人系统技术》
2022年第2期71-79,共9页
Unmanned Systems Technology
基金
国家自然科学基金(61876187)。
关键词
机动目标跟踪
径向基神经网络
多模型
特征向量
目标状态估计
广义回归神经网络
鲁棒性
Maneuvering Target Tracking
Radial-based Neural Network
Multi-Model
Feature Vector
Target State Estimate
Generalized Regression Neural Network
Robustness