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. 过去、现在以及未来：朝着增强感知能力的方向探索同步定位与建图[J]. IEEE Transactions on robotics, 2016, 32(6):1309-1332.

文献内容 ：
SLAM经典综述

阅读笔记：

过去、现在与未来：同时定位与 Mapping heading towards the era of robust perception

I. INTRODUCTION

Two issues (which frequently engage in debates when taking place at robotics events)

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

• 可靠性能
• 高级认知能力
• 资源敏感性
• 任务驱动的推理

Paper organization

II. ANATOMY OF A MODERN SLAM SYSTEM

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

The fundamental realization driving contemporary SLAM systems lies in the fact that the Jacobian matrix exhibits sparsity characteristics, which stem from the structural properties inherent to these factor graphs.

Sensor-dependent SLAM front-end

The preprocessing carried out in the frontend is dependent on sensors, due to the concept of features varying based on which data stream we consider.

III. LONG-TERM AUTONOMY I: ROBUSTNESS

Open Problems

• fail-safe SLAM as well as recovery mechanisms
• robustness against hardware failure
• metrical relocalization technique
• time-varying and deformable map representations
• self-tuning parameter adjustment system

IV. LONG-TERM AUTONOMY II: SCALABILITY

two ways to reduce the complexity of factor graph optimization

sparse approximation techniques by trading off information loss while balancing the trade-off between memory requirements and computational efficiency

* Node and edge sparsification
* Continuous-time trajectory estimation

out-of-core and multi-robot methods which partition the computational tasks across multiple robots and processors

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

Open Problems

• Mapping upkeep
  • Distributed mapping with resilience
  • Learning, forgetting, and remembering
  • Systems constrained by resource limitations

V. REPRESENTATION I: METRIC REASONING

the categories of 3D metric representation

• 基于关键点的稀疏表示
• 低层次原始密集表示
• 边界和空间划分密集表示
• 高层次物体导向表示

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

primitive method

feature-based approach

disadvantage

• 基于特征类型的依赖
• 对多种检测与匹配阈值的依赖
• 必须采用鲁棒的估计技术以应对错误对应的问题
• 多数特征检测器中优先优化速度而非精确度

dense, direct methods

merit

The proposed method demonstrates superior performance compared to existing feature-based techniques in challenging textures where focus is lacking and motion-induced blur is present.

shortcoming

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

improvement

Semi-dense methods

This approach effectively reduces the computation burden associated with conventional dense methods by focusing on pixels characterized by strong gradient magnitudes.

semi-direct methods

Instead, we can employ both sparse features (including corners and edges) and direct methods.

These methods utilize sparse feature representations, enabling the simultaneous estimation of both structure and motion.

Open Problems

• High-order and powerful representations are integral to the success of SLAM systems.
• The pursuit of optimal representations is a cornerstone of effective system design.
• Self-adaptive and dynamic representations enable systems to autonomously adjust their operation parameters.

VI. REPRESENTATION II: SEMANTIC REASONING

Semantic SLAM

semantic parsing at the fundamental layer can be modeled as a classification task.

three main ways to deal with semantic mapping

• SLAM facilitates Semantics learning.
• Semantics learning is facilitated by SLAM.
• The integration of SLAM and Semematics inference enables comprehensive understanding.

Open Problems

• Semantic mapping goes beyond mere categorization problems.
  • Ignorance, acuteness of awareness, and adjustment.
  • Reasoning based on semantics is another important aspect.

VII. NEW THEORETICAL TOOLS FOR SLAM

optimization approaches

EKF

factor graph

advantage

  * accuracy, efficiency
  * an elegant framework

Open Problems

• Generality, assures of benefits, and validation are essential elements.
• Is it weak or strong duality?
• Robustness against outliers is a critical feature.

VIII. ACTIVE SLAM

detail

conception

The challenge to manage robot's motion in order to reduce the uncertainty of its mapping and localization is typically referred to as active SLAM.

A popular framework

• Choosing perspectives for analysis
  • Calculating the value derived from performing each action.
  • Engaging in executing actions and ceasing further exploration.

Open Problems

• Rapid and precise forecasting of future state evolutions
• It’s time to consider ending active SLAM: When should you cease this activity?
• assurances on system performance.

IX. NEW SENSORS AND OTHER FRONTIERS

New and Unconventional Sensors for SLAM

• 双目相机和结构光相机
• 距离测量相机
• 光场相机
• 基于事件的相机
• What these sensors are used for performing SLAM has not been considered by them.

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！