有效点较少的动态场景下单目视觉SLAM算法  

Monocular vision SLAM algorithm with fewer effective points under dynamic scenes

作　　者：孙平[1] 闫冬 SUN Ping;YAN Dong(School of Artificial Intelligence,Shenyang University of Technology,Shenyang 110870,China)

机构地区：[1]沈阳工业大学人工智能学院,沈阳110870

出　　处：《沈阳工业大学学报》2022年第4期431-437,共7页Journal of Shenyang University of Technology

基　　金：国家重点研发计划项目(2016YFD0700104);辽宁省自然科学基金项目(2019ZD0203).

摘　　要：为了解决有效点较少的动态复杂场景下视觉SLAM准确定位问题,提出了一种基于自适应RANSAC动态特征点剔除的单目视觉SLAM算法.通过ARANSAC算法估计图像间的透视变换矩阵,并扭曲上一帧获得一个估计图像,使上一帧中的点转换到当前帧的坐标系下.通过计算特征点在估计图像和当前帧的光流值,区分并剔除ORB-SLAM2中的动态特征点,从而消除动态物体对SLAM定位性能的影响.利用TUM数据集的动态序列对本文算法进行仿真,并与ORB-SLAM2算法进行对比.结果表明,视觉SLAM算法绝对轨迹误差的标准偏差降低84.00%~96.11%,平移和旋转漂移的标准偏差最佳效果分别降低94.00%和96.44%,明显减少了视觉SLAM算法位姿估计的误差.本文算法能够在有效点较少的动态场景下,消除动态物体对视觉SLAM定位性能的影响,提高定位精度.In order to solve the problem for accurate localization by using visual SLAM under complex dynamic scenes with fewer effective points,a monocular visual SLAM algorithm was proposed.The perspective transformation matrix between images was estimated by ARANSAC algorithm,and the dynamic feature points in ORB-SLAM2 were distinguished and eliminated by calculating the optical flow values of feature points in estimated images and current frames.Using the dynamic sequence of TUM data set,the as-proposed algorithm was simulated and compared with the ORB-SLAM2 algorithm.The results show that the standard deviation of the absolute trajectory error of as-proposed visual SLAM algorithm reduces by 84.00%to 96.11%,and the optimal standard deviation of translation and rotation drift reduces by 94.00%and 96.44%,respectively,significantly lowering the error of pose estimation by visual SLAM algorithm.The as-proposed algorithm can eliminate the influence of dynamic objects on the localization performance of visual SLAM under dynamic scenes with fewer effective points to improve the localization accuracy.

关 键 词：单目 视觉SLAM 动态场景 ARANSAC算法 光流 透视变换 动态特征点剔除 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]