论文部分内容阅读
为了实时得到搭载双目相机的机器人准确的3维位姿状态和环境信息,提出一种融合直接法与特征法的双目SLAM(同时定位与地图创建)算法.该算法主要分为4个线程:跟踪线程、特征提取线程、局部建图线程和闭环线程.跟踪线程通过最小化图像光度误差,获取双目的初始位姿估计和特征对应关系,而后通过最小化局部地图点的重投影误差,得到更为准确的机器人位姿估计.特征提取线程负责提取关键帧的关键点和描述子,能够保证待处理的关键帧较多时不影响后续局部建图线程的执行.局部建图线程管理局部地图,执行局部BA(光束平差法),优化局部关键帧位姿和局部地图点的位置,提高SLAM的局部一致性.闭环线程通过对关键帧的闭环检测和优化,提高SLAM的全局一致性.另外利用闭环线程处理机器人被绑架后重回已探测环境的定位问题.KITTI数据集、TUM数据集以及采集的双目数据实验表明,本文算法相对于ORB-SLAM2算法,在保证定位精度的同时,有效提高了相机位姿的输出帧率,并且在机器人被绑架的情况下,能够得到更为丰富的姿态信息和环境信息.
In order to obtain accurate 3D attitude and environment information of robots equipped with binocular cameras in real time, a binocular SLAM (Simultaneous Localization and Map Creation) algorithm combining both direct method and feature method is proposed.The algorithm is mainly divided into four threads : Tracking thread, feature extraction thread, local drawing thread and closed loop thread.The tracking thread obtains the initial pose estimation and feature correspondence of two eyes by minimizing the photometric error of the image, and then, by minimizing the reprojection error of the local map point, Get a more accurate robot pose estimation.Feature extraction thread is responsible for extracting the key points and descriptors of the keyframe, to ensure that the keyframe to be processed does not affect the subsequent execution of the local assembly thread.Local drawing thread management part Map, the implementation of partial BA (beam adjustment method), to optimize the location of local keyframe pose and local map points to improve the SLAM local consistency. Closed-loop threads through the closed-loop detection and optimization of key frames to improve SLAM global consistency In addition, closed-loop threading is used to deal with the positioning problem of robots being abducted and returned to the probed environment.KITTI dataset, TUM dataset and double The experimental results show that compared with the ORB-SLAM2 algorithm, the proposed algorithm can effectively improve the position and frame output of the camera while ensuring the positioning accuracy. In addition, the robotic system can obtain richer attitude information and environment information.