SLAM algorithms are grouped into three categories depending on the techniques used; filter-based SLAM, graph SLAM, and Deep Learning (DL) based SLAM. The filter-based SLAM treats the problem as a state estimation problem. The state, which comprises the pose and map of the robot, is updated by a filter iteratively based on measurements as the robot explores. The DL-based SLAM solves the problem by replacing an entire end-to-end process with DL networks. Only sub tasks in the graph SLAM can be replaced with DL networks. However, it is reasonable to classify such algorithms as the graph SLAM. The graph SLAM is currently the state-of-the-art algorithm. The end-to-end DL-based SLAM does have recent promising results but is not mature, and the filter-based SLAM performs worse than the graph SLAM in general. Therefore, this paper limits the discussion to the graph SLAM. SLAM algorithms can be classified further based on the primary sensor, for example, visual SLAM, LiDAR (Light Detection and Ranging) SLAM, and so forth. Inertial Measurement Units (IMU) or Inertial Navigation Systems (INS) are often used together with the primary sensor to improve the accuracy of pose estimation.