LVI-SAM: Tightly-coupled Lidar-Visual-Inertial Odometryvia Smoothing and Mapping
该两个子系统以紧耦合的方式设计。对于VIS初始化,LIS提供初始参数(系统状态x和IMU零偏b且假设b不变)。VIS进行视觉特征跟踪时,引入激光雷达帧可提供特征深度。LIS在进行扫描匹配时的初始猜测是来自优化视觉重投影和IMU测量误差,此作为因子图的一个约束。IMU对点云去畸变后,LIS提取边和面特征并与滑动窗口中维护的特征地图进行匹配。LIS得到的系统估计再发给VIS以方便VIS初始化。闭环矫正是VIS识别候选匹配,LIS再优化,其中视觉里程计、激光雷达里程计、IMU预积分和闭环约束在因子图中进行优化。最后利用优化后的IMU零偏传播测量值来进行位置估计。
主要贡献:
- 1、A tightly-coupled LVIO framework built atop a factor graph, which achieves both multi-sensor fusion and global optimization aided by place recognition.
- 2、Our framework bypasses failed sub-systems via failure detection, making it robust to sensor degradation.
- 3、Our framework is extensively validated with data gathered across varied scales, platforms, and environments.