融合直接法与特征法的快速双目SLAM算法

发布时间：2019-03-31 20:30

【摘要】：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.
[Abstract]:In order to obtain accurate 3-D pose state and environment information of robot with binocular camera in real time, a binocular SLAM (simultaneous location and map creation) algorithm is proposed, which combines direct method and feature method. The algorithm is divided into four threads: trace thread, feature extraction thread, local drawing thread and closed loop thread. By minimizing the luminosity error of the image, the tracking thread can obtain the relationship between the initial position and pose estimation and the feature correspondence of the binocular. Then, by minimizing the re-projection error of the local map point, a more accurate pose estimation of the robot can be obtained. The feature extraction thread is responsible for extracting the key points and descriptors of the key frames, which ensures that the execution of subsequent local drawing threads is not affected by the large number of key frames to be processed. Local mapping thread manages local map, executes local BA (beam adjustment method), optimizes the position and position of local key frame and local map point, and improves the local consistency of SLAM. The closed loop thread improves the global consistency of SLAM by detecting and optimizing the key frames. In addition, closed-loop thread processing is used to deal with the localization problem of robot after being kidnapped and returned to the detected environment. Experiments on KITTI data set, TUM data set and binocular data acquisition show that compared with the ORB-SLAM2 algorithm, the proposed algorithm guarantees the positioning accuracy at the same time. It improves the output frame rate of camera pose effectively, and can obtain more information of attitude and environment when the robot is kidnapped.
【作者单位】：火箭军工程大学控制工程系;成都信息工程大学控制工程学院;
【分类号】：TP242

【相似文献】