一种结合光流追踪的双目视觉里程计算法

A stereo visual odometry algorithm combined with optical flow tracking

摘要:

针对特征点法视觉里程计频繁计算和匹配描述子导致系统实时性能变差的问题,提出一种结合光流追踪的双目视觉里程计算法;首先进行初始化,生成初始的关键帧和地图点,随后在追踪线程中使用光流追踪特征点获取匹配关系,计算并优化相机位姿;满足生成关键帧的条件后,将当前帧设置为关键帧,提取图像的ORB特征点,并使用描述子匹配获取与上一关键帧特征点的匹配关系,三角化生成新的地图点;最后在优化线程中将新的关键帧和地图点使用滑动窗口算法进行优化,剔除冗余关键帧和地图点;在KITTI数据集上的实验结果表明,所提出的算法轨迹误差与双目模式下的ORB-SLAM3算法处于同一水平,同时实时性能有大幅度提高,追踪每一帧图像的平均时间从52ms降至16ms,在保证高精度的情况下运行速度大大提高,具有较高的实用价值.

Abstract:

Aiming at the problem that the real-time performance of the system deteriorates due to the frequent calculation and matching descriptors of the feature-based visual odometry, a stereo visual odometry combined with optical flow tracking is proposed. Firstly, initialization is performed to generate initial key frames and map points, then in the tracking thread, use the optical flow tracking feature points to obtain the matching relationship, calculate and optimize the camera pose. After meeting the conditions for generating key frames, set the current frame as a key frame, extract the ORB feature points of the image, use descriptor matching to obtain the matching relationship with the feature points of the previous key frame, and triangulate to generate new map points. Secondly , the new keyframes and map points are optimized using the sliding window algorithm in the optimization thread,and redundant keyframes and map points are eliminated. In the KITTI data the experimental results show that the trajectory error of the proposed algorithm is at the same level as the ORB-SLAM3 algorithm in stereo mode, and the real-time performance is greatly improved.The average time for tracking each frame of image is reduced from 52ms to 16ms. In the case of ensuring high precision, the running speed is greatly improved, and it has high practical value.