计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

基于区域分布概率密度估计的轨迹分类方法

曹卫权,李智翔,魏强,褚衍杰   

  1. (盲信号处理重点实验室,成都 610041)
  • 收稿日期:2017-02-14 出版日期:2018-04-15 发布日期:2018-04-15
  • 作者简介:曹卫权(1989—),男,博士,主研方向为机器学习、数据挖掘;李智翔、魏强,博士;褚衍杰,工程师、博士。

Trajectory Classification Method Based on Probability Density Estimation of Regional Distribution

CAO Weiquan,LI Zhixiang,WEI Qiang,CHU Yanjie   

  1. (National Key Laboratory of Science and Technology on Blind Signal Processing,Chengdu 610041,China)
  • Received:2017-02-14 Online:2018-04-15 Published:2018-04-15

摘要: 区域分布是运动目标的重要特征,可用于目标轨迹分类。已有分类方法往往假设轨迹片段呈矩形簇状或混合高斯状分布,限制了轨迹分类精度的提升。为此,提出一种基于核密度估计和最大似然判决的轨迹分类方法,消除已有分类方法对数据分布模型的先验假设,进而解决因模型不适配导致的轨迹分类准确率受限问题。实验结果表明,相较于最小描述长度划分、高斯混合模型等方法,该方法对参数不敏感,训练时间明显缩短,轨迹分类准确率提升5%~15%。

关键词: 轨迹分类, 最小描述长度, 核密度估计, 高斯核, 最大似然

Abstract: Regional distribution is an important feature of moving target,which can be used for target trajectory classification.Previous methods assumed that segments of trajectories distribute as rectangles or Gaussian components,thus the accuracy of trajectory classification is significantly limited.To solve the problem,a novel method that takes use of kernel density estimation and maximum likelihood principle is proposed.It has no assumption on the data,so the problem of trajectory classification accuracy limited by model discomfort is solved.The experimental results show that compared with the existing Minimum Description Length(MDL) partition and Gaussian Mixture Model(GMM),the proposed method is insensitive to parameters,training time is shortened,and track classification accuracy is increased by 5%~15%.

Key words: trajectory classification, Minimum Description Length(MDL), kernel density estimation, Gaussian kernel, maximum likelihood

中图分类号: