SLAM论文精读系列：（第一篇）Past, present, and future of simultaneous localization and mapping

• 本篇博文由 南湖游子 原创，转载请注明出处！

文献来源：

Cadena C, Carlone L, Carrillo H, et al. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age[J]. IEEE Transactions on robotics, 2016, 32(6):1309-1332.

文献内容 ：
SLAM经典综述

阅读笔记：

Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age

I. INTRODUCTION

two questions (that often animate discussions during robotics conferences)

• Do autonomous robots really need SLAM?

  • visual-inertial navigation (VIN)
  • loop closures
  • globally consistent map

• Is SLAM solved?

  • evaluation aspects of maturity of SLAM

    • robot
    • environment
    • performance requirements

key requirements of the robust-perception age for SLAM

• robust performance
• high-level understanding
• resource awareness
• task-driven inference

Paper organization

II. ANATOMY OF A MODERN SLAM SYSTEM

maximum-a-posteriori estimation and the SLAM back-end

• The key insight behind modern SLAM solvers is that the Jacobian matrix is sparse and its sparsity structure is dictated by the topology of the underlying factor graph.

Sensor-dependent SLAM front-end

• The pre-processing that happens in the front-end is sensor dependent, since the notion of feature changes depending on the input data stream we consider.

III. LONG-TERM AUTONOMY I: ROBUSTNESS

Open Problems

• failsafe SLAM and recovery
• robustness to hardware failure
• metric relocalization
• time varying and deformable maps
• automatic parameter tuning

IV. LONG-TERM AUTONOMY II: SCALABILITY

two ways to reduce the complexity of factor graph optimization

• sparsification methods which trade off information loss for memory and computational efficiency

  • Node and edge sparsification
  • Continuous-time trajectory estimation

• out-of-core and multi-robot methods which split the computation among many robots/processors

  • Out-of-core (parallel) SLAM
  • Distributed multi robot SLAM

Open Problems

• Map maintenance
• Robust distributed mapping
• Learning, forgetting, remembering
• Resource-constrained platforms

V. REPRESENTATION I: METRIC REASONING

the categories of 3D metric representation

• Landmark-based sparse representations
• Low-level raw dense representations
• Boundary and spatial-partitioning dense representations
• High-level object-based representations

Which one is best: feature-based or dense, direct methods?

• primitive method

  • feature-based approach

    • disadvantage

      • dependence on feature type
      • reliance on numerous detection and matching thresholds
      • necessity for robust estimation techniques to deal with incorrect correspondences
      • optimization for speed rather than precision in most feature detectors

  • dense, direct methods

    • merit

      • outperforming feature-based methods in scenes with poor texture, defocus, and motion blur

    • shortcoming

      • requiring high computing power (GPUs) for real-time performance

• improvement

  • Semi-dense methods

    • overcoming the high-computation requirement of dense method by exploiting only pixels with strong gradients

  • semi-direct methods

    • instead leverage both sparse features (such as corners or edges) and direct methods

  • both relying on sparse features and allowing joint estimation of structure and motion

Open Problems

• High-level, expressive representations in SLAM
• Optimal Representations
• Automatic, Adaptive Representations

VI. REPRESENTATION II: SEMANTIC REASONING

Semantic SLAM

• semantic parsing at the basic level being formulated as a classification problem

• three main ways to deal with semantic mapping

  • SLAM helps Semantics
  • Semantics helps SLAM
  • Joint SLAM and Semantics inference

Open Problems

• Semantic mapping is much more than a categorization problem
• Ignorance, awareness, and adaptation
• Semantic-based reasoning

VII. NEW THEORETICAL TOOLS FOR SLAM

optimization approaches

• EKF

• factor graph

  • advantage

    • accuracy, efficiency
    • an elegant framework

Open Problems

• Generality, guarantees, verification
• Weak or Strong duality?
• Resilience to outliers

VIII. ACTIVE SLAM

detail

• conception

  • The problem of controlling robot’s motion in order to minimize the uncertainty of its map representation and localization is usually named active SLAM.

• A popular framework

  • Selecting vantage points
  • Computing the utility of an action
  • Executing actions or terminating exploration

Open Problems

• Fast and accurate predictions of future states
• Enough is enough: When do you stop doing active SLAM?
• Performance guarantees

IX. NEW SENSORS AND OTHER FRONTIERS

New and Unconventional Sensors for SLAM

• Stereo and structure-light cameras
• Range cameras
• Light-field cameras
• Event-based cameras
• Using haptic sensors to perform SLAM having not been considered

New Frontiers: Deep Learning

• Open Problems

  • Perceptual tool
  • Practical deployment
  • Bootstrapping

X. CONCLUSION

• 本篇博文由 南湖游子 原创，转载请注明出处！