计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (3): 171-178.DOI: 10.3778/j.issn.1002-8331.1711-0022

• 图形图像处理 • 上一篇    下一篇

改进协方差矩阵的智能车视觉目标跟踪方法

刘红星1,2,胡广地1,朱晓媛1,李进龙3   

  1. 1.西南交通大学 汽车研究院,成都 610031
    2.西南交通大学 电气工程学院,成都 610031
    3.西南交通大学 交通运输与物流学院,成都 610031
  • 出版日期:2019-02-01 发布日期:2019-01-24

Vision Target Tracking Method of Intelligent Vehicle Based on Improved Covariance Matrices

LIU Hongxing1,2, HU Guangdi1, ZHU Xiaoyuan1, LI Jinlong3   

  1. 1.Automotive Research Institute, Southwest Jiaotong University, Chengdu 610031, China
    2.School of Electrical Engineering, Southwest Jiaotong University, Chengdu 610031, China
    3.School of Transportation and Logistics, Southwest Jiaotong University, Chengdu 610031, China
  • Online:2019-02-01 Published:2019-01-24

摘要: 智能车辆视觉目标具有非线性、噪声分布非高斯性的典型特点,现有算法难以实时估计目标的状态。针对识别物体复杂且多变,很难用完全的特征来描述待识别目标及其背景的不断变化,提出了一种用于融合颜色特征及SURF(Speed-Up Robust Features)特征的协方差矩阵来改进粒子滤波算法,从而提升视觉目标跟踪的实时性,满足智能车辆的要求。首先,对采集的图像进行预处理来获取感兴趣区域。接着,通过融合颜色特征及SURF特征构造范围感兴趣区域(Region Of Interest,ROI)的目标特征协方差矩阵,建立目标状态预测模型及状态观测模型,用于改进粒子滤波算法粒子重采样过程,实现对目标的精确跟踪。最后,将该方法与Mean-shift算法和颜色属性(CN)算法进行对比。实验结果表明,在智能车视觉跟踪过程中对光环境瞬时变化、目标物体存在短时遮挡以及目标物体姿态改变时,该算法在满足智能车辆对实时性要求的前提下,有效提升算法的精确度及鲁棒性。

关键词: 视觉目标追踪, 粒子滤波算法, 协方差矩阵, 特征融合

Abstract: There are the salient features of target nonlinearity and non-Gauss distribution of noise with the vision target recognition method of intelligent vehicle, so it’s hard for existing algorithms to estimate target state in real time. Due to the complexity and changeability of objects needed to be recognized, it’s almost impossible to adopt complete features to describe the target and its dynamic background. A covariance matrix fused with color features and speed-up robust features is proposed in this paper, which is used for particle filtering algorithm, thus achieving the accurate tracking of target. Firstly, the collected image is pretreated to obtain the region of interest. Secondly, a target feature covariance matrix in the ROI is constructed by fusing color features and speed-up robust features. Then, target state prediction model and state observation model used for particle resampling process in improved particle filter algorithm are built, which can implement accurate tracking of targets. Finally, the method is compared with traditional particle filter method characterized by single color features and speed-up robust features. Test results show that, for vision target recognition and tracking of intelligent vehicle when light environment is instantaneous changed, target object has short duration occlusion or target object changes attitude, the accuracy and robustness of the algorithm are effectively improved with the premise of meeting real-time requirement.

Key words: target recognition and tracking, particle filtering algorithm, covariance matrices, feature fusion