基于ORBSLAM2改进的单线程双目SLAM系统  

Single-thread stereo SLAM system based on ORBSLAM2

作　　者：王劭靖 WANG Shaojing(School of Information,North China University of Technology,Beijing 100144,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《智能计算机与应用》2023年第1期84-90,99,共8页Intelligent Computer and Applications

摘　　要：ORBSLAM2是一种稳定、鲁棒的实时SLAM系统,但其四线程结构给CPU带来了较重的负担,很难移植到嵌入式设备中。本文针对此问题对ORBSLAM2进行3处改进,将平滑的运动约束思想应用到运动模型的加速度中,使用加速运动模型预测相机的位姿,并添加环形匹配策略来剔除特征点的误匹配,快速、有效地确定良好的相机位姿初值和数据关联关系;将车联网领域中轨迹压缩的思想应用到SLAM中,提出一种基于开放窗口的关键帧聚类算法对位姿图中的位姿节点进行聚类,减少位姿图优化的规模;使用基于拉普拉斯分布的解耦合优化三维空间点、相机的朝向与平移的方法代替局部捆绑调整(BA),提高优化的精度和速度。在KITTI和New Tsukuba数据集上的实验结果表明,改进后的ORBSLAM2能够以单线程的形式实时运行并且准确地估计相机的位姿。ORBSLAM2 is a stable and robust real-time SLAM system, but its four-thread structure puts a heavy burden on the CPU and is difficult to be ported to embedded devices. This paper addresses this problem by making three improvements to ORBSLAM2, applying the idea of smooth motion constraints to the acceleration of the motion model, using the accelerated motion model to predict the camera’s poses, and adding a ring matching strategy to eliminate the mis-matching of feature points to quickly and efficiently determine good initial values of camera poses and data association relationships;applying the idea of trajectory compression in the field of vehicle networking to SLAM. A key frame clustering algorithm based on open window is proposed to cluster the positional nodes in the positional map to reduce the scale of positional map optimization;a method based on the decoupling of Laplace distribution to optimize the orientation and translation of 3D spatial points and cameras is used instead of local bundle adjustment(BA) to improve the accuracy and speed of optimization. Experimental results on the KITTI and New Tsukuba datasets show that the improved ORBSLAM2 can run in real time as a single thread and accurately estimate the camera’s poses.

关 键 词：ORBSLAM2 开放窗口 单线程 拉普拉斯分布 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]