摘要
针对动态背景下视觉跟踪存在的错误率大、鲁棒性不强以及多目相机信息融合难等问题,提出一种基于双目匹配的视觉跟踪算法.结合双目相机的物理结构特点,该算法以四边形闭环方法对特征点进行匹配,实现特征点的双目匹配和立体跟踪,并优化匹配的搜索结构.算法首先采用高斯-拉普拉斯模板对图像进行滤波,利用Harris角点检测算法提取特征点并依次构造描述子序列,然后以绝对误差和作为匹配代价的衡量准则,建立四边搜索准则进行邻近搜索,同时引入RANSAC(随机采样一致性)算法进行可靠性筛选,最终通过4点构成的四边形闭环检测实现了特征点跟踪精度的改进.通过对不同分辨率、不同路况图像集进行测试实验,所提出的四边形跟踪算法特征跟踪正确率可达99.80%,鲁棒性和精度均优于光流法.
A visual tracking algorithm based on binocular matching is proposed to solve the problems such as high error rate, low robustness and poor information fusion existing in current visual tracking methods under dynamic background. Based on the physical structure of binocular camera, the proposed algorithm adopts a closed quadrilateral method to match feature points so as to implement the binocular matching and 3D tracking and optimize the searching structure. The binocular images are filtered by Gauss-Laplace template, and the feature points extracted by the Harris corner detection algorithm are encapsulated orderly into feature descriptor. Then, the sum of absolute difference is used as the matching criteria, a 4-side searching criteria is set up for neighbor searching, and RANSAC （random sample consensus） algorithm is introduced for reliability screening. Finally, the accuracy of feature tracking is improved through the 4 points formed quadrilateral closed-loop detection. To evaluate the performance of the proposed method, images with different resolutions under different road conditions are collected, and test experiments are conducted. The experimental results show that the average tracking precision of the proposed method reaches 99.80%. The robustness and tracking accuracy of the proposed method is superior to the optical flow method.
出处
《机器人》
EI
CSCD
北大核心
2015年第6期674-682,共9页
Robot
基金
国家自然科学基金(61305111
91120307
91320301)
安徽省自然科学基金(1508085MF133)
关键词
视觉跟踪
信息融合
双目匹配
四边形闭环
visual tracking
information fusion
binocular matching
closed quadrilateral
