基于改进LK光流的视觉SLAM算法  

Visual SLAM Algorithm Based on Improved LK Optical Flow

作　　者：余功兴 王宪伦[1] 秦晓 孙志鹏 YU Gong-xing;WANG Xian-lun;QIN Xiao;SUN Zhi-peng(College of Electromechanical Engineering,Qingdao University of Science and Technology,Qingdao 266061,China)

机构地区：[1]青岛科技大学机电工程学院,青岛266061

出　　处：《组合机床与自动化加工技术》2023年第3期35-38,42,共5页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金资助项目(51105213);山东省重点研发计划(2018GGX103015)。

摘　　要：针对基于点特征的视觉SLAM(同步定位与建图)实时性差的问题,以ORB-SLAM3框架为基础,用自监督深度神经网络替换其点特征提取器,并提出一种基于改进LK光流的关键点跟踪匹配方法。首先,根据匀加速运动模型预测初始关键点在当前帧的对应关系及坐标,并通过基于金字塔的LK光流跟踪算法建立鲁棒的关键点匹配关系;然后,基于网格运动统计(GMS)及自适应局部仿射匹配(AdaLAM)算法剔除误匹配点对。在公开数据集上的实验表明,在不损失匹配精度的前提下,改进算法每帧图像的平均处理时间比ORB-SLAM3减少约58.06%,有效提高了系统的实时性和鲁棒性。To address the problem of poor real-time performance of point feature-based visual SLAM(simultaneous localization and map building),we replace its point feature extractor with a self-supervised deep neural network based on the ORB-SLAM3 framework and propose a key point tracking and matching method based on improved LK optical flow.First,the correspondence and coordinates of the initial key points in the current frame are predicted based on the uniform acceleration motion model,and the robust key point matching relationship is established by the pyramid-based LK optical flow tracking algorithm;then,the mis-matched point pairs are rejected based on the grid motion statistics(GMS)and adaptive local affine matching(AdaLAM)algorithm.Experiments on publicly available datasets show that the average processing time per image frame of the improved algorithm is reduced by about 58.06%compared with ORB-SLAM3 without losing matching accuracy,effectively improving the real-time and robustness of the system.

关 键 词：LK光流 视觉SLAM GCNv2 实时性 匀加速运动模型 

分 类 号：TH165[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]