08_MMS_Systems.md

MMS Systems

1- SLAM integrated mobile mapping system in complex urban environments

• They establish a SLAM-integrated MMS with the capability of efficient, consistent and robust mapping in complex en- vironments where GNSS signal is intermittently lost.
• Fisrt: derived the probabilistic model for SLAM-integrated MMS, which abstractly proved the advances of adding the feedback from the mapping module to the localization module.
• Second: Describe details of the hierarchical factor graph optimization structure, which is the core of the proposed method.
  • Essential factor graph: dumps all redundant factors to keep the global graph concise and tight.
  • iSAM (incremental smoothing and mapping) algorithm is employed off-line to efficiently solve the global optimization.
• Point cloud registration is performed on GPU using surfel-based approach to speed up.

Traditional MMS

1. combining INS and GNSS
2. +Lidar
3. GNSS-denied area:
   • automatic registration of point cloud map from multiple passes: requires data from multiple passes of the same scene and require manual adjustment.
   • Hussnain et al. (2018) extracted the known landmarks in the environment as control points. (against SLAM idea)

Maximum a posteriori estimation

• multi-source data fusion technique
• From the perspective of probability, the estimation of robot poses can be transformed as maximizing the conditional distribution of robot pose, given the noise model of all on- board sensors.
  • (1) filtering methods, e.g. Extended Kalman Filter (EKF), and Particle Filter (PF),
  • (2) and smoothing methods.
    • Full smoothing: optimizes entire history of states
    • Fixed-lag smoothing: optimizes states falling within a given time window
      • compromise of filter approaches and full smoothing approaches. accuracy is higher than filter approaches because they re-linearize past measurements.)
      • The disadvantage is that the marginalization of old tates destructs the sparsity of hessian matrix, which decreases the effciency of optimization.
    • iSAM (Incremental Smoothing and Mappng): STOA. Opimizes a small subset of variables dentified as affected nodes by the new measurement

LiDAR-based SLAM

Failure: structureless environment or under rapid rotation

• 1. classical point-based ICP registration
  • Costs too much on closest points searching
  • May get stuck into the local optimum, when a bad initial is provided
• 2. LOAM
  • divided SLAM prblem into the odometry and the mapping thread (performed feature-based scan-to-scan registration on odometry at high frequency, while preformed feature-based scan-to-map registration on the parallel Mapping thread at lower frequency)
  • lacks an optimization back-end, which makes error accumulation
• 3. Lego_Loam: Aded an optimization back-end for LOAM\
• 4. SuMa(Surfel Mapping): surfel-based SLAM
  • A global surfel-based map was constructed and maintained.
  • The surfel-based ICP algorithm is a variation of classical ICP.
  • It project points as the depth image, so that correspondences can be directly found by indexing the same pixel in a pair of successive depth images, which avoids the costly searching and increase the convergency of the estimation iteration.
  • he changes of poses were estimated by exploiting the projective data associations between current laser scan and a rendered model view from the global point cloud map.
• 5. SuMa++: SuMa + RangeNet++
  • Improve the registration accuracy in dynamic environment: Remove the moving objects in the point cloud and build a semantic map.

SLAM + MMS (during GNSS outages)

• Indoortechniques: indoor environment is well-structured, i.e. indoor free space is flat and full of lines and surfaces: Good for 2D SLAM solutions

  • Gmapping
  • Cartographer

• Totally GNSS denied environment

• GNSS is sporadic:
  GNSS signals are unavailable: Blocked or effected by multi-path effect
  They set the time periods of GNSS-denied as 5s, 10s, 15s, 20s, 25s, 30s, 40s, 50s and 60s

  • 1- Pseudo-GNSS/INS system
    These sensors provide the data stream for map construction.
    The key of this is transforming the pose estimation by SLAM to Pseudo-GNSS signals. To improve the efficiency of the mapping optimization
    EKF

    GNSS/INS-based MMS	proposed Pseudo-GNSS/INS framework

  • 2- Image-assisted gnss/ins navigation for uav-based mobile mapping systems during gnss outages.
    Hierarchical filter along with smoothers
    1- The first stage is a tightly coupled Kalman Filter (KF) followed by a smoother using GNSS/INS measurements as inputs.
    2- The second stage is a loosely coupled KF utilizing the output from the first stage and a vision-based trajectory to aid trajectory refinement during GNSS outages.

    idea	detailed

System overview

1- Green blocks: raw data

• IMU/GNSS > EKF
• GNSS Unavailable: IMU only
• Lidar: Relative transformation

2- Pose Graph(first)

• IMU preintegrated factors (GNSS signalk locked)
  • Or odometry factors: GNSS/IMU filter
• scan-to-map registration factors (surfel-based:SuMa)
• small loop factors (surfel-based:SuMa)

3- when GNSS signal is lost, construction time of a submap span a long period of time: surfel-based approach may fail.

• change to use point-based registration

  

Hierarchical optimization structure

Local refinement	Global refinement
Scan-to-map registration factor	Map-to-map registration factor

Cited

1- Continuous-Time Laser Frames Associating and Mapping via Multilayer Optimization (MDPI sensors 2020)

• No GNSS/INS
• Normal distributions transform (NDT) and ICP
• Continuous-time laser frames associating and mapping framework
  • 1- Intraframe point cloud segmentation (RANSAC)
  • 2- Interframe point cloud association (NDT)
    match the consecutive frames with the ground points and effective non-ground points
  • 3- Submap matching (ICP)
    the first frame of the submap is used as a key frame
  • 4- Closed loop detection (NDT)
  • 5- Graph optimization:
• Three-layer point cloud association technique.

🌕 Note: (LOAM):Uses features to align consecutive frames
NDT:Normal Distributions Transform :interframe association of sparse point clouds (match the point clouds with a global map)
The interframe point cloud association method:
1- ICP: easy to fall into the local optimum
2- fast point feature histogram (FPFH): expensive computational costs
3- Super-4PCS: global search with (RANSAC), ostly in computation but has higher association accuracy
4- normal distributions transform (NDT): fastest and relatively reliable method

2- MULLS: Versatile LiDAR SLAM via Multi-metric Linear Least Square(ICRA 2021)

• With low drift
• Real-time performance
• Rank 13th KITTI benchmark
• Geometric feature points extraction
  • ground point G, facade F, roof R, pillar P, beam B, and vertex V
• Multi-metric linear least square ICP: (multiple distance metrics)

Overall workflow of MULLS-SLAM

Front-end:

• Roughly classified feature points
• Multi-metric Linear Square (MULLS) ICP
• Local map updating

Back-end:

• Submap-based loop closure
• TEASER-based global registration
• Inter & inner submap PGO

Point Cloud Registration (scan matching):

• Local: Requires good initial guess to finely align two overlapping scans without stuck in the local minima
• Global: Coarsely align them 
• Deep Learning
  • First learn to embed points 
  • Learn to match keypoints 
  • Optimization PointNet, DGCNN, PointNetLK, DCP,…

Local Matching:

• ICP (Iterative closest point)
• Sparse point-based:
  • LOAM (The edge and planar points)
  • LeGO-LOAM (Conduct ground segmentation)
• Dense projective normal ICP:
  • Suma
  • SuMa++ (Semantic masks)
• Limitations:
  • Requires an initial guess
  • Lose 3D information due to the operation based on range image or scan-line

Pose Graph Optimization

Pipeline of geometric feature points extraction and encoding	Pipeline of the multi-metric linear least square ICP