摘要
随着自动驾驶、智能交通的发展,跟踪算法成为热点问题。本文主要针对的是在毫米波雷达路侧交通场景中通过对广义标签多伯努利滤波器(Generalized Labeled Multi-Bernoulli Filter,GLMB)参数估计从而保证滤波器在时变交通监测中目标跟踪的性能。参数选择是制约滤波器性能的主要因素之一,掌握其特性,具有十分重要的意义。传统的跟踪滤波会在特定的场景中使用一套固定的参数,当场景变化时,滤波器参数无法及时调整,导致跟踪性能降低。针对该问题,本文将长短期记忆神经网络(Long Short Term Memory network,LSTM)引入GLMB滤波器参数估计领域,通过雷达数据训练神经网络,使其具备对滤波器参数估计能力。使用毫米波雷达数据构建的数据集训练神经网络,训练完成后将使用测试数据集验证神经网络对参数估计结果。不同交通场景的雷达实测数据验证结果表明,与人为设定的固定参数方法相比,该方法可以使滤波器在时变交通监测中及时对参数进行估计与调优,提升了GLMB滤波器目标跟踪的性能。
With the development of autonomous driving and intelligent transportation,tracking algorithms have become a hot issue.This paper is mainly aimed at predicting the parameters of the Generalized Labeled Multi-Bernoulli Filter(GLMB)in the Millimeter wave radar roadside traffic scene to ensure the performance of the filter in the time-varying traffic monitoring.Parameter selection is one of the main factors that restrict the performance of the filter,and it is of great significance to master its characteristics.Traditional tracking filtering uses a fixed set of parameters in a specific scene.When the scene changes,the filter parameters cannot be tuned in time,resulting in reduced tracking performance.In response to this problem,this paper introduces the Long Short Term Memory network(LSTM)into the field of GLMB filter parameter estimation,and trains the neural network through data to enable it to estimate the filter parameters.Use Millimeter wave radar data to construct a data set to train the neural network.After the training is completed,the test data will be used to obtain the neural network′s parameter prediction results.The verification results of radar measured data in different traffic scenarios show that compared with the artificially set fixed parameter method,this method can enable the filter to estimate and tune the parameters in time in timevarying traffic monitoring,and improve the GLMB filter target tracking Performance.
作者
李洋
王铜
王彦平
林赟
申文杰
LI Yang;WANG Tong;WANG Yanping;LIN Yun;SHEN Wenjie(Radar Monitoring Technology Laboratory,School of Information Science and Technology,North China University of Technology,Beijing 100144,China)
出处
《信号处理》
CSCD
北大核心
2022年第1期148-156,共9页
Journal of Signal Processing
基金
国家自然科学基金面上项目(61971456)
国家自然科学基金重点国际合作研究项目(61860206013)
国家重点研发计划资助(2018YFC1505103)。
关键词
神经网络
参数估计
长短期记忆神经网络
多目标跟踪
广义标签多伯努利滤波
neural networks
parameter estimation
long short term memory network
multi-target tracking
Generalized labeled multi-Bernoulli filter
