计算机工程 ›› 2020, Vol. 46 ›› Issue (7): 50-57.doi: 10.19678/j.issn.1000-3428.0055074

• 人工智能与模式识别 • 上一篇    下一篇

基于高斯混合-变分自编码器的轨迹预测算法

张显炀, 朱晓宇, 林浩申, 刘刚, 安喜彬   

  1. 火箭军工程大学 核工程学院, 西安 710025
  • 收稿日期:2019-05-30 修回日期:2019-07-10 发布日期:2019-07-23
  • 作者简介:张显炀(1995-),男,硕士研究生,主研方向为机器学习、行为预测;朱晓宇,硕士;林浩申,博士;刘刚,教授、博士;安喜彬,博士。
  • 基金项目:
    国防科技"引领"基金(18-163-00-75-004-078-01)。

Trajectory Prediction Algorithm Based on Gaussian Mixture-Variational Autoencoder

ZHANG Xianyang, ZHU Xiaoyu, LIN Haoshen, LIU Gang, AN Xibin   

  1. School of Nuclear Engineering, Rocket Force University of Engineering, Xi'an 710025, China
  • Received:2019-05-30 Revised:2019-07-10 Published:2019-07-23

摘要: 海面舰船的轨迹预测对预测精度和实时性具有较高要求,而舰船轨迹数据特征的高复杂度特性,导致传统预测算法精度低、耗时长,难以达到良好的预测效果。为此,提出一种基于变分自编码器的海面舰船轨迹预测算法。将轨迹坐标数据集转化为轨迹移动矢量集,使用变分自编码器完成轨迹运动特征的提取与生成预测。同时为提高轨迹预测精度,将变分自编码网络的隐空间分布设定为混合高斯分布,使其更符合真实的数据分布特征,并在隐空间完成轨迹特征的分类,实现端到端的轨迹预测。仿真结果表明,相较于传统预测算法GMMTP和VAETP,该算法的预测误差分别降低了85.48%和35.59%。

关键词: 轨迹预测, 变分自编码器, 混合高斯模型, 无监督学习, 端到端学习

Abstract: The trajectory prediction of warships requires high accuracy and real-time performance,but the high complexity of trajectory data features of warships causes the traditional prediction algorithms to be inaccurate and time-consuming,reducing prediction performance.To address the problem,this paper proposes a warship trajectory prediction algorithm based on Variational Autoencoder(VAE).The trajectory coordinate data set is transformed into a trajectory motion vector set,and the trajectory motion features are extracted and generated by using variational autoencoder.Also,in order to improve the prediction accuracy,the hidden space distribution of the variational autoencoding network is set to be mixture Gaussian distribution,which is closer to the features of real data distribution.Then the classification of trajectory features is accomplished in hidden space to implement end-to-end trajectory prediction.Simulation results show that compared with the traditional trajectory prediction algorithms,GMMTP and VAETP,the proposed algorithm can reduce the prediction error rate by 85.48% and 35.59% respectively.

Key words: trajectory prediction, Variational Autoencoder(VAE), Gaussian Mixture Model(GMM), unsupervised learning, end-to-end learning

中图分类号: