摘要
多传感器建图与定位SLAM系统(simultaneous localization and mapping)在室外长距离跨度环境中,由于各传感器信息融合不正确、特征匹配错误,或传感器状态信息不可信,导致建图精度不足,轨迹漂移甚至建图崩溃。对此,提出一种基于因子图优化的多传感器信息紧耦合算法(tightly-coupled lidar-visual-inertial odometry via smoothing,mapping and DBSCAN,LVI-SMAD),将前端点云和视觉信息联合的聚类结果作为因子图优化约束,以一种较低帧的约束形式加入到较高帧的点云地图输出中,加强了点云与视觉信息的紧耦合,解决了激光雷达与相机间信息匹配错误的问题,同时将该约束作为某一传感器信息不可信时的约束补充,减小了传感器信息不稳定情况下的定位漂移,提高了算法一致性。实验证明,在低坡度长跨度的工作环境中,LVI-SMAD与LVI-SAM对比,绝对轨迹误差降低了39.90%,与LIO-SAM对比降低了63.09%;在高坡度工作环境中,与LVI-SAM对比,绝对轨迹误差减少41.08%,与LIO-SAM对比减少64.87%,证明了算法的有效性与可行性。
In outdoor long-range and large-scale environments,multi-sensor simultaneous localization and mapping(SLAM)systems often encounter challenges related to insufficient mapping accuracy,drifting,and map crashes.These issues arise due to incorrect sensor fusion,errors in feature matching,and unreliable sensor state information.To mitigate these problems,this paper proposes a factor graph optimization-based algorithm called LVI-SMAD(tightly-coupled lidar-visual-inertial odometry via smoothing,mapping and DBSCAN)that addresses tightly-coupled multi-sensor information.It utilizes clustering results obtained from the joint front-end point cloud and visual information as constraints for factor graph optimization.These constraints are incorporated into the output point cloud map of higher frames,enhancing the cohesive relationship between point cloud and visual data.This approach overcomes the inaccurate information matching between lidar and camera.Moreover,the constraints act as supplements when certain sensor information becomes unreliable,thereby reducing positioning drift in scenarios with unstable sensor data and enhancing algorithm consistency.Our experimental results demonstrate the significant performance improvement of LVI-SMAD compared with existing methods.In low-slope and long-span environments,LVI-SMAD reduces absolute trajectory error by 39.90%compared with LVI-SAM and by 63.09%compared with LIO-SAM.Similarly,in high-slope mapping environments,LVI-SMAD achieves a substantial reduction of 41.08%in absolute trajectory error compared to LVI-SAM and a 64.87%reduction compared to LIO-SAM.These results provide compelling evidence for the effectiveness and feasibility of the proposed algorithm.
作者
甄子杰
汪汗青
王诚
霍文渊
赵毅
ZHEN Zijie;WANG Hanqing;WANG Cheng;HUO Wenyuan;ZHAO Yi(School of Electrical and Information Engineering,Yunnan Minzu University,Kunming 650000,China)
出处
《重庆理工大学学报（自然科学）》
CAS
北大核心
2024年第3期156-165,共10页
Journal of Chongqing University of Technology：Natural Science
基金
国家自然科学基金项目(61963038)。
关键词
多传感器融合
树优化密度聚类
前端数据处理
因子图优化
multi-sensor fusion
density clustering with tree
front-end data processing
factor graph optimization
