SLAM survey

참고자료

- https://arxiv.org/abs/2108.01654

- https://www.koreascience.or.kr/article/JAKO201411560021024.pdf

 

 

 

Sparse visual SLAM

• MonoSLAM (2007)
  • Andrew J. Davison et al.
  • Speed up SfM algorithm using probabilistic filters
  • Can obtain landmarks and trajectory of the camera in real-time
• PTAM (2007)
  • Georg Klein, David Murray
  • Single thread에서 sequential 하게 동작하던 MonoSLAM을 개선
  • Bundle adjustment가 filter-based approach 보다 우월하다는 것을 보여줌

• ORB-SLAM (2015)
  • Raul Mur-Artal et al.
  • Bag of Words
  • ORB feature
  • g2o optimizer
• ORB-SLAM2 (2017)
  • Raul Mur-Artal et al.
  • Supports stereo and RGB-D cameras
• VINS (2017a)
  • Tong Qin et al.
  • Use IMU sensors
  • Kalman filter
• VINS-Mono (2018)
  • Tong Qin et al.
• SOFT-SLAM (2018)
  • Igor Cvisic et al. 
  • Improvement of ORB-SLAM2
  • Completely deterministic (i.e. always return same result for same input)
  • SOFT keypoints are not invariant with respect to rotation
• ORB-SLAM3 (2020)
  • Visual inertial approach
  • Updated place recognition module
  • Support for pinhole and fisheye camera models.
  • Allows localization in a multi-map setup.

 

 

 

Dense visual SLAM

• DTAM (2011b)
  • Richard A. Newcombe et al.
  • Photometric error minimization
  • Generated dense 3D map for each pixel
• KineticFusion (2011a)
  • Richard A. Newcombe et al.
  • Used depth camera to build a dense 3D map.
  • Used TSDF (truncated signed distance function) to describe each pixel.
  • ICP (iterative closest point) algorithm to map each depth image to a map.
• ElasticFusion (2016)
  • Thomas Whelan et al.
  • Used direct representation of the surfaces of RGB-D cameras
  • Used non-rigid deformation model for data fusion.
  • No global optimization of the graph of camera positions
  • Most of the steps take place on GPU...
• SVO (2013)
  • Jakob Engel et al.
  • Real time on a robot processor (CPU)
  • Does not use every pixels in the image, but only those that have a neglibigle gradient.
• LSD-SLAM (2014)
  • Jakob Engel et al.
  • Used sim(3) metric to solve a problem with an uncertain scale
  • Used probabilistic inference to determine the error in constructing 3D maps.
  • Optimization using g2o library
• DSO (2017)
  • Jakob Engel et al.
  • Photometric error + geometric error
  • Used probabilistic pixel sampling instead of assuming smoothness.
• LSDO (2018)
  • Xiang Gao et al.
  • Added graph optimization to get a complete solution to the visual SLAM problem.
  • Actually a combination of ORB-SLAM and LSD-SLAM approaches.
• DSO with Rolling Shutter (2019)
  • David Schubert et al.
  • Modification of DSO (2017)
  • Process frame obtained from floating shutter.
  • IMU data are used as additional constraints for the optimization problem.

 

 

 

Dynamic visual SLAM

• DynaSLAM (2018)
  • Bescos et al.
  • Possible to detect a priory dytnamic objets.
  • Possible to detect objects that may be static at the moment, but the dynamic in essence.
  • Pixel-wise semantic segmentation of potentially moveable objects using Mask R-CNN.
  • Background inpainting by using information from previous views.
• DynSLAM (2018)
  • Barsan et al.
  • Large scale dynamic environments.
  • Takes stereo images and compute depth map (using ELAS or DispNet) and sparse scene flow (libviso2)
  • Multi-task Network Cascade to fild dynamic objects.
  • Final result is a static map without dynamic object with help of InfiniTAM.
• DRE-SLAM (2019)
  • Yang et al.
  • Uses 2 wheel encoders.
  • Uses ORB features from RGB-D camera.
  • YOLOv3 for dynamic object detection
  • Loop closure with BoW
  • OctoMap.

 

 

Visual-Inertial Odometry

• VINS-Mono (2011)
  • Maddern et al.
  • tightely coupled visual inertial system.
  • IMU preintegration between camera frames.
  • Optimization using Ceres solver
  • Loop closure
• OpenVINS (2020)
  • Geneva et al.
  • Multi-State Kalman Filter (MSCKF) based VIO estimator.
• VINS-Mono (2018)
  • Qin et al.
• VINS-Fusion (2019)
  • Qin et al.
  • Extension of VINS-Mono.
  • Supports multiple visual-inertial sensor types.
  • Supports global sensors (GPS, Barometer)
  • Graph optimization module
  • Loop closure
• Basalt (2019)
  • Usenko et al.
  • Graph-based VIO apporach.
  • KLT feature tracking and Gauss-Newton non-linerar optimization
• Kimera (2020)
  • Antoni et al.
  • Open source c++ library
  • VIO module with GTSAM-based VIO approach.
  • Robust position graph optimizer (RPGO) for global trajectory estimation.
  • Lightweight 3D mesh module (Kimera-Mesher) for fast 3D mesh reconstruction.
  • Obstacle avoidance and a dense 3D semantic reconstruction module (Kimera semantics).
  • Semantically annotates the 3D mesh using 2D pixel-wise semantic segmentation based on deep learning.
  • 개쩐다...