水下视觉SLAM分段式光束平差算法  

Segmented bundle adjustment algorithm for underwater vision SLAM

作　　者：白云鹏 徐会希[1,2,3] 吕凤天 BAI Yunpeng;XU Huixi;L Fengtian(State Key Laboratory of Robotics,Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China;Institute for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang 110169,China;Key Laboratory of Marine Robotics,Liaoning Province,Shenyang 110169,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院沈阳自动化研究所机器人学国家重点实验室,辽宁沈阳110016 [2]中国科学院机器人与智能制造创新研究院,辽宁沈阳110169 [3]辽宁省水下机器人重点实验室,辽宁沈阳110169 [4]中国科学院大学,北京100049

出　　处：《测绘通报》2024年第11期7-12,共6页Bulletin of Surveying and Mapping

基　　金：中科院战略性先导科技专项(A类)(XDA22040103);辽宁省自然科学基金计划面上项目(2022-MS-035)。

摘　　要：自主水下机器人(AUV)采用视觉SLAM系统可以实现近距离精确定位,但是面对水下大规模场景时,后端优化采用的光束平差算法(BA)存在内存不足和计算效率低的问题,对此本文提出了一种改进的分段式BA优化算法。首先采用基于运动模式的分段方法,根据相机的直行运动和转弯运动将轨迹分段后对每个子分段分别进行BA优化;然后各子分段采用动态调整优化权重的求解方法,根据不同子分段的运动模式动态调整优化参数;最后针对BA代价函数的求解,采用改进的列文伯格-马夸尔特(L-M)求解算法,将信赖域定义为可调参数,优化由雅克比矩阵非正定性引发的算法稳定性问题,提高运算效率。在数据集上的试验结果表明,在运行时间较长、环境较恶劣的数据集序列上,本文算法相较于ORB-SLAM3算法有更好的精度,同时全局BA的效率有显著提高。Autonomous underwater vehicle(AUV)can achieve close-range accurate positioning by using visual SLAM system,but when facing large-scale underwater scenes,the back-end optimization using bundle adjustment(BA)algorithm has the problems of insufficient memory and low computational efficiency.To solve these problems,an improved segmented BA optimization algorithm is proposed.A segmentation method based on motion pattern is used to segment the trajectory according to the straight motion and turning motion of the camera,and then BA optimization is performed on each sub-segment respectively.Each sub-segment is solved by dynamically adjusting the optimization weight,and the optimization parameters are dynamically adjusted according to the motion patterns of different sub-segments.For the solving of BA cost function,the improved Levenberg-Marquadt(L-M)algorithm is adopted,the trust region is defined as the tunable parameter,which reduces the non-convergence problem caused by the singularity of the Jacobian matrix and improves the operation efficiency.According to the experimental results on the dataset,the proposed algorithm has better accuracy than the ORB-SLAM3 algorithm when it runs for a long time and the environment is harsh,and the efficiency of the global BA is significantly improved.

关 键 词：水下视觉SLAM 后端优化 光束平差法 分段式BA L-M算法 

分 类 号：P237[天文地球—摄影测量与遥感]