【免费】基于监督学习的在线过滤器用于使用雷达测量进行目标跟踪1资源-CSDN文库

需积分: 0 124 浏览量 2022-08-04 11:56:34 上传评论收藏 1.65MB PDF 举报

"基于监督学习的在线过滤器用于使用雷达测量进行目标跟踪" 本文介绍了一种基于监督学习的在线过滤器，用于使用雷达测量进行目标跟踪。该过滤器可以在未知和复杂环境中实现高精度的目标状态估计，克服了传统贝叶斯过滤器在动态模型匹配度和先验信息方面的限制。一、监督学习在目标跟踪中的应用监督学习是机器学习的一个分支，通过学习标记数据来训练模型，以便对未知数据进行预测。在目标跟踪领域中，监督学习可以用于建立雷达测量与目标状态之间的映射关系，从而实现目标状态的估计。二、在线过滤器的原理在线过滤器是一种实时处理雷达测量数据的滤波器，能够实时地估计目标状态。该过滤器的关键是建立雷达测量与目标状态之间的映射关系，以便根据雷达测量数据来估计目标状态。三、基于监督学习的在线过滤器本文提出的基于监督学习的在线过滤器首先建立雷达测量与目标状态之间的映射关系，然后使用数据驱动的方式来实现目标状态的估计。该过滤器不需要先验信息，继承了高精度的估计能力在未知和复杂环境中。四、雷达测量数据处理雷达测量数据处理是目标跟踪中的关键步骤。该步骤包括数据预处理、特征提取、映射关系建立等。基于监督学习的在线过滤器可以对雷达测量数据进行实时处理，从而实现目标状态的估计。五、XGBoost算法在目标跟踪中的应用 XGBoost算法是一种基于梯度boosting的机器学习算法，能够对复杂数据进行高精度的预测。在目标跟踪领域中，XGBoost算法可以用于建立雷达测量与目标状态之间的映射关系，从而实现目标状态的估计。六、实验结果实验结果表明，基于监督学习的在线过滤器可以在未知和复杂环境中实现高精度的目标状态估计，超过传统贝叶斯过滤器的估计精度。七、结论本文提出了一种基于监督学习的在线过滤器，用于使用雷达测量进行目标跟踪。该过滤器可以在未知和复杂环境中实现高精度的目标状态估计，克服了传统贝叶斯过滤器在动态模型匹配度和先验信息方面的限制。该技术有广泛的应用前景，在雷达目标跟踪、机器人 navigation、自动驾驶等领域中具有重要的价值。

资源详情

资源评论

资源推荐

Supervised Learning Based Online Filters for

Targets Tracking Using Radar Measurements

Jie Deng, Wei Yi, Kai Zeng, Qiyun Peng and Xiaobo Yang

University of Electronic Science and Technology of China, Chengdu, P.R. China

E-mail: jiedeng sc@outlook.com; kussoyi@gmail.com; 201921010930@std.uestc.edu.cn; qiyunpeng13@std.uestc.edu.cn

Abstract—In the ﬁeld of the radar target tracking, the state

ﬁltering plays an important role in estimating the target state.

One of the widely-adopted ﬁlter is the Bayesian ﬁlter, which

requires the prior information and an accurate modeling of the

real tracking scene. Thus, the matching degree of the dynamic

model has a key impact on the state estimation accuracy of

the Bayesian ﬁlter. However, the target dynamic and radar

measurement models cannot be approximated perfectly in an

unknown and complicated environment, and the state estimation

accuracy of the model-based ﬁlter is limited. To address the

limitations of the Bayesian ﬁlter, a supervised learning based

online ﬁlter for target tracking is proposed in this paper. In the

proposed ﬁlter, a mapping among the radar measurements is ﬁrst

established in the context of the polar coordinate system. Then,

based on data-driven, the state ﬁltering is directly implemented to

obtain the state estimate by using this mapping relationship. As

such, the prior information is not required in the proposed ﬁlter,

hence the proposed ﬁlter inherits a good estimation accuracy

in unknown and complicated environments. Finally, simulation

experiments clarify the effectiveness of our proposed ﬁlter via

comparing with the traditional ﬁlter.

Index Terms—XGBoost, radar measurements, data-driven,

ﬁlter.

I. INTRODUCTION

In the application of radar target tracking, the state ﬁlter

is an important module. It is mainly used to address the

state estimation problem in time series by using the historical

noise measurement data. Herein, the Bayesian ﬁlter is a

classic ﬁltering framework [1], which is widely used in the

target tracking. Nowadays, based on this framework, a series

of available ﬁlters have been developed. For example, the

Kalman ﬁlter (KF), the extended Kalman ﬁlter (EKF), the

unscented Kalman ﬁlter (UKF) and the particle ﬁlter (PF).

More speciﬁcally, for the radar target tracking problems, the

EKF, the UKF and the PF are often used to solve the ﬁltering

problem in nonlinear non-Gaussian systems [1], [2].

For the traditional Bayesian ﬁlter, it is necessary to model

the dynamic system as well as set and estimate reasonable

prior parameters to achieve an accurate description of the sys-

tem (i.e., the model and the real system are required to match

as much as possible). However, the target statistical character-

istics cannot be obtained perfectly in practical environments,

which makes the estimation of the priori information incorrect.

This work was supported in part by the National Natural Science Foundation

of China under Grants 61771110, in part by Chang Jiang Scholars Program,

in part by the 111 Project B17008.

(Corresponding author: Wei Yi, E-mail: kussoyi@gmail.com.)

Therefore, the mismatch of the preset parameter model will

cause the tracking performance to decline or even diverge in

these cases, which is also a limitation of the model-based ﬁlter.

In order to address this problem, many works were performed.

For example, Dang et al. [3] and Tripathi et al. [4] proposed

an adaptive ﬁlter for maneuvering targets and unknown noise.

An adaptive scheme is proposed in [5] by considering the

measurement noise covariance distribution. Although these

studies have effects on some speciﬁc scenarios, they are not a

general method and also increase the complexity of the model,

which motivates us to consider a new solution way.

Supervised learning is the machine learning task of learning

a function that maps an input to an output based on example

input-output pairs [6]. It can avoid modeling the motion system

by building a data mapping relationship and get rid of the lim-

itations of model-based ﬁlters. Nowadays, supervised learning

has been widely used in data prediction, image recognition

and text classiﬁcation, which has achieved good results [7].

There are still some studies on the application of supervised

learning in the ﬁlters. For example, the combination of the

neural network (NN) and KF was ﬁrst proposed in [8], in

which the NN is used to train the residual. Guo et al. [9]

used the trained NN to compensate the EKF drift in the GPS

system. However, they both used supervised learning to train

and assist the residuals in [8] and [9]. Subsequently, a random

forest (RF) based ﬁlter that directly maps the measurement

to the estimation was proposed in [10]. On the basis of [10],

using the eXtreme Gradient Boosting (XGBoost) to improve

the estimation accuracy in [11]. Gao et al. [12] considered

using Recurrent Neural Network (RNN) for nonlinear target

tracking, which has a good estimation ability. However, none

of the above-mentioned work [10] – [12] extracts the target

motion characteristics, directly used the existing data for

training. Therefore, consider extracting the motion change

information of training data was proposed in [13], including

three aspects of time, space, and angle, then they proposed a

XGBoost based ﬁlter (XGBF). However, only the Cartesian

coordinate system is considered in [13], the polar coordinate

system of the radar are not analyzed. If the training data is

directly converted from the polar coordinate system to the

Cartesian coordinate system for training in [13], there will

be inﬂuences in the near and far regions and conversion error.

The nonlinear conversion of the data will lead to inconsistent

training characteristics in different regions, which will affect

the training effect.

2020 IEEE Radar Conference (RadarConf20)

Authorized licensed use limited to: Nanjing Univ of Post & Telecommunications. Downloaded on February 19,2021 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.

本内容试读结束，登录后可阅读更多

下载后可阅读完整内容，剩余5页未读，立即下载